Cover page

Image Name: BSG-IDE

BSG-IDE is a cross platform scientiﬁc presentation, preparation and distribution tool developed at airis4D. It is

an open source Creative Commons licenced software developed in Python with over 10,000 lines of code and

nearly three months of developer time.

Managing Editor Chief Editor Editorial Board Correspondence

Ninan Sajeeth Philip Abraham Mulamoottil K Babu Joseph The Chief Editor

Ajit K Kembhavi airis4D

Geetha Paul Thelliyoor - 689544

Arun Kumar Aniyan India

Sindhu G

Journal Publisher Details

Publisher : airis4D, Thelliyoor 689544, India

Website : www.airis4d.com

Email : nsp@airis4d.com

Phone : +919497552476

Editorial

by Fr Dr Abraham Mulamoottil

airis4D, Vol.2, No.12, 2024

www.airis4d.com

This edition also starts with the picture gallery

of ”Igniting Young Minds” highlights the eﬀorts of

airis4D in promoting scientiﬁc education and collab-

oration. Resource persons Geetha Paul and Ninan

Sajeeth Philip led educational programs at institutions

like the Breakthrough Science Society in Kottayam,

Marthoma College in Thiruvalla, and Bethany Academy

in Vennikulam, engaging young students in science and

research. The organization also signed a Memorandum

of Understanding (MoU) with Marthoma College for

joint research programs, strengthening academic part-

nerships. Additionally, Professors Naresh Dadhich and

Ajit Kembhavi, former directors of the Inter-University

Centre for Astronomy and Astrophysics (IUCAA) in

Pune, visited airis4D, furthering its connections with

renowned institutions in astronomy and astrophysics.

These initiatives showcase airis4D’s commitment to

advancing scientiﬁc education and fostering collabora-

tion among academic communities.

The articles start with “ LVM Architectures” by

Blesson George. Large Vision Models (LVMs) mark a

signiﬁcant leap in computer vision by leveraging exten-

sive datasets, advanced architectures, and scalable de-

signs. Key innovations include the Swin Transformer,

which addresses the limitations of Vision Transform-

ers through local window-based attention, hierarchical

representation, and eﬃcient handling of high-resolution

tasks like image classiﬁcation and semantic segmen-

tation. Similarly, DETR (DEtection TRansformer)

revolutionizes object detection by simplifying tradi-

tional pipelines with an end-to-end transformer-based

approach and bipartite matching for precise and ef-

ﬁcient predictions. Together, these models enhance

scalability, eﬃciency, and versatility, showcasing the

transformative potential of transformer-based architec-

tures in advancing visual understanding.

The article by Jinsu Ann Mathew “Domain Gener-

ation Algorithms” (DGAs) are sophisticated tools used

by malware to evade detection and maintain communi-

cation with command-and-control (C2) servers. They

generate dynamic and unpredictable domain names,

often using a combination of seed values, domain gen-

eration logic, and entropy sources. Seed values en-

sure synchronization between malware and C2 servers,

while logic and entropy introduce variability and ran-

domness. By generating domains in high volume and

at frequent intervals, DGAs complicate defenders’ ef-

forts to block all possible connections. Understanding

these components is crucial for developing eﬀective

cybersecurity strategies against DGA-driven threats.

”How to Improve Your ML Models” by Linn

Abraham says: that improving machine learning (ML)

models centres on enhancing their ability to generalize

from training data to real-world scenarios. Eﬀective

generalization depends more on the structure of the

data than the model itself. Steps like cleaning data, en-

gineering informative features, and reducing noise are

essential before considering model adjustments. Key

strategies for avoiding overﬁtting and enhancing gen-

eralization include early stopping, regularization tech-

niques (e.g., weight penalties and dropout), and resiz-

ing model parameters. These practices align with the

”manifold hypothesis,” which suggests that natural data

lies on low-dimensional, structured subspaces within

high-dimensional input spaces, enabling ML models

to interpolate eﬀectively within these regions.

The article, Black Hole Stories-14: Gravitational

Wave Emission, provides an overview of electromag-

netic wave properties and their relevance to understand-

ing gravitational waves. Electromagnetic Waves are

generated by oscillating electric and magnetic ﬁelds

and are described by Maxwell’s equations. They prop-

agate at the speed of light and include phenomena like

radiation emitted by accelerated charges, with patterns

depending on charge velocity and acceleration. Radi-

ation Properties: Accelerated charges emit energy as

radiation, with emission patterns inﬂuenced by their

velocity and acceleration. Relativistic eﬀects intensify

forward-directed radiation. Electromagnetic waves in

a vacuum follow sinusoidal patterns, with energy ﬂow

deﬁned by electric and magnetic ﬁelds. In a medium,

wave properties change based on the medium’s charac-

teristics. Electromagnetic waves can exhibit linear, cir-

cular, or elliptical polarisation depending on the phase

and amplitude of electric ﬁeld components. The ar-

ticle by Professor Ajit Kembhavi sets the stage for

future discussions on gravitational waves, noting that

understanding electromagnetic waves helps grasp the

complexities of gravitational phenomena governed by

Einstein’s general relativity.

The article by Sindhu G “Unveiling the Stars: A

Deep Dive into Stellar Spectral Classiﬁcation explores

the history, methodology, and applications of classi-

fying stars based on their light spectra. It traces the

evolution from Angelo Secchi’s early spectral classes

to the Harvard and Yerkes (MK) systems, which group

stars by temperature, luminosity, and spectral features.

Specialized categories like Wolf-Rayet stars and L, T,

Y dwarfs expand the classiﬁcation. Modern advances,

including automated surveys and high-resolution spec-

troscopy, enhance our understanding of stellar evo-

lution, galactic structure, and exoplanets, showcasing

spectral classiﬁcation as a cornerstone of astrophysical

research.

Innovating Autism Support: AI-Driven Approaches

for a More Inclusive Society by Kalyani Bagri explores

the transformative potential of Artiﬁcial Intelligence

(AI) in addressing Autism Spectrum Disorders (ASD).

It begins by highlighting ASD’s multifactorial causes,

including genetic, environmental, and neurobiologi-

cal factors, and underscores the importance of early

diagnosis and intervention to mitigate symptoms and

support developmental outcomes. AI-driven tools are

revolutionizing ASD care in several ways:Diagnosis:

Natural Language Processing (NLP) enables early and

accurate identiﬁcation of communication and social

deﬁcits. Therapies: AI platforms provide personal-

ized therapy plans, virtual therapists, and social skill

training tailored to individual needs. Management:

Predictive analytics and mobile apps empower care-

givers with insights and resources for ongoing support.

In India, AI innovations are bridging gaps in ASD care

by oﬀering personalized e-learning platforms, assis-

tive communication technologies, and developmental

gaming. Initiatives like NGO-led training programs

and corporate inclusion eﬀorts promote employment

for adults with ASD, fostering autonomy and societal

acceptance. The article concludes that leveraging tech-

nology and fostering collaboration can create a more

inclusive society for individuals with ASD while ad-

dressing challenges like bias, data privacy, and acces-

sibility.

The article by Ninan Sajeeth Philip “”BSG-IDE:

Revolutionizing Academic Presentations with AI-Powered

Automation and Real-Time Media Integration” intro-

duces BSG-IDE, an innovative open-source tool for

creating academic presentations, developed by Ninan

Sajeeth Philip under airis4D. Unlike traditional LaTeX-

based tools, BSG-IDE integrates artiﬁcial intelligence,

real-time media processing, and smart content manage-

ment to streamline slide creation. Its user-friendly de-

sign bridges the gap between expertise in a subject and

the eﬀort needed to prepare engaging slides. Key fea-

tures include Dynamic Layouts: Automatically adjust

slide content for optimal presentation. Media Integra-

tion: AI-powered suggestions for images, videos, and

animations with citation support. Real-Time Syntax

Feedback: Highlights commands and media elements,

aiding error detection. Dual-Screen Mode: Displays

speaker notes and slides simultaneously. Overleaf

Export: Facilitates cloud-based LaTeX collaboration.

Available as a Python package, BSG-IDE prioritizes

simplicity and innovation, making it an invaluable tool

for researchers and educators seeking visually appeal-

iii

ing, mathematically rigorous presentations.

Igniting Young Minds

by News Desk

airis4D, Vol.2, No.12, 2024

www.airis4d.com

Geetha Paul and Ninan Sajeeth Philip were resource persons at Breakthrough Science Society program at

Kottayam and on a seminar at Marthoma College, Thiruvalla and Bethany Academy, Vennikulam. The airis4D

has signed an MoU for joint research programs with Marhoma College.

vii

Professors Naresh Dadhich and Professor Ajit Kembhavi spent three days as guests at airis4D with family.

They both were former directors of IUCAA, Pune, the prestigious Inter University Centre for Astronomy and

Astrophysics.

viii

Contents

Editorial ii

Igniting Young Minds vi

I Artiﬁcial Intelligence and Machine Learning 1

1 LVM Architectures 2

1.1 SWIN Transformer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2

1.2 DETR Transformer . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3

1.3 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4

2 Domain Generation Algorithm 5

2.1 Seed Value . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5

2.2 Domain Generation Logic . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2.3 Entropy Source . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6

2.4 Frequency and Volume of Domain Generation: . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

2.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7

3 How to Improve Your ML models 9

3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

3.2 Generalization in ML models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9

3.3 How to Improve Generalization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

3.4 Regularization Techniques . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10

II Astronomy and Astrophysics 12

1 Black Hole Stories-14

Gravitational Wave Emission 13

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

1.2 Electromagnetic Waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13

1.3 Radiation from an Accelerated Charge . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14

1.4 Plane Electromagnetic Waves . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 15

1.5 Polarisation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 16

2 Unveiling the Stars: A Deep Dive into Stellar Spectral Classiﬁcation 18

2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2.2 Historical Development of Stellar Spectral Classiﬁcation . . . . . . . . . . . . . . . . . . . . . . 18

2.3 The Basics of Stellar Spectral Classiﬁcation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18

2.4 The Harvard Spectral Classiﬁcation System . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2.5 Yerkes Spectral Classiﬁcation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 19

2.6 Specialized Stellar Spectral Classes . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

CONTENTS

2.7 The Physics of Stellar Spectra . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

2.8 Applications of Spectral Classiﬁcation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 20

2.9 Advances in Spectral Classiﬁcation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

2.10 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21

III Biosciences 22

1 Innovating Autism Support: AI-Driven Approaches for a More Inclusive Society 23

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

1.2 Understanding the Causes of ASD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 23

1.3 Importance of Early Diagnosis and Intervention . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24

1.4 Transforming Autism Diagnosis and Care with AI . . . . . . . . . . . . . . . . . . . . . . . . . . 24

1.5 Leveraging Technology for Individuals with ASD in India . . . . . . . . . . . . . . . . . . . . . . 25

1.6 Technology-Driven Solutions for ASD . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 25

1.7 Empowering Employment: Adult Autism Workstations . . . . . . . . . . . . . . . . . . . . . . . 26

IV General 28

1 BSG-IDE: Revolutionizing Academic Presentations with AI-Powered Automation and

Real-Time Media Integration 29

1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

1.2 Simplicity is the Key . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

1.3 Installation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29

1.4 Let’s Get Started . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

1.5 New Presentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 30

1.6 Pympress . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

1.7 Overleaf Integration . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

1.8 The Magic World of BSG-IDE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 31

Part I

Artiﬁcial Intelligence and Machine Learning

LVM Architectures

by Blesson George

airis4D, Vol.2, No.12, 2024

www.airis4d.com

Large Vision Models (LVMs) represent a sig-

niﬁcant leap in the ﬁeld of computer vision by lever-

aging extensive datasets, advanced architectures, and

powerful computational resources. These models are

designed to understand, process, and analyze visual

data with remarkable precision and scalability. LVMs

often employ self-supervised learning techniques to

learn rich visual representations, making them adapt-

able to a wide range of downstream tasks, such as

object detection, image segmentation, and scene un-

derstanding. By scaling both model size and train-

ing data, LVMs achieve state-of-the-art performance

across various benchmarks, pushing the boundaries of

visual comprehension in artiﬁcial intelligence.

Vision Transformers (ViT): Vision Transform-

ers (ViTs) have revolutionized the approach to pro-

cessing visual data by introducing transformer-based

architectures, which were originally successful in nat-

ural language processing, into the domain of computer

vision. Unlike convolutional neural networks (CNNs),

ViTs divide images into patches and treat them as se-

quences, applying attention mechanisms to learn global

and contextual features eﬃciently. This method has

proven highly eﬀective, particularly for tasks requiring

a comprehensive understanding of images. ViTs ex-

cel in handling large-scale datasets and have become

a cornerstone of modern large vision models, enabling

enhanced interpretability and robustness in various ap-

plications.

Please refer to the previous issue of the airis4D

Journal for a detailed exploration of the Vision Trans-

former (ViT) architecture. In this session, we will focus

on discussing two other signiﬁcant networks in the ﬁeld

of computer vision. These networks have contributed

substantially to advancements in processing and under-

standing visual data, oﬀering unique approaches and

beneﬁts. Let us delve into their architectures, applica-

tions, and how they complement or extend the capabil-

ities of ViT and other contemporary models.

1.1 SWIN Transformer

The Swin Transformer represents a signiﬁcant ad-

vancement in the ﬁeld of computer vision. Building on

the success of Vision Transformers (ViTs), it addresses

key challenges associated with processing large-scale

visual data and paves the way for transformer-based

models to be used as general-purpose backbones for

computer vision tasks.

1.1.1 Challenges in Vision Transformers

(ViTs)

While ViTs revolutionized image processing by

treating image patches as tokens (akin to words in text),

they face several limitations:

Scale Variability: Images contain visual ele-

ments that vary widely in scale, which ViTs

struggle to model eﬀectively.

Computational Complexity: The global self-

attention mechanism in ViTs scales quadratically

with image size, making it computationally ex-

pensive for high-resolution tasks.

Single-Scale Representation: ViTs generate single-

resolution feature maps, which are unsuitable for

dense prediction tasks like object detection and

semantic segmentation.

1.2 DETR Transformer

Figure 1: The Swin Transformer creates a structured

way of understanding images by building hierarchical

feature maps. This is done by combining smaller parts

of the image (patches, shown in gray) as the model

goes deeper. It simpliﬁes the calculations needed by

focusing on smaller regions (local windows, shown in

red) instead of analyzing the whole image at once. This

approach makes it much faster and more eﬃcient, es-

pecially for large images. As a result, the Swin Trans-

former can be used for a wide variety of tasks, like

recognizing objects in images or understanding their

details.

In contrast, earlier Vision Transformers worked with a

single low-resolution version of the image and required

signiﬁcantly more calculations because they analyzed

the whole image at once. This made them slower and

less practical for tasks involving high-resolution im-

ages.

1.1.2 The Swin Transformer: A Solution

The Swin Transformer introduces innovative so-

lutions to these challenges through:

1. Local Window-Based Attention: Self-attention

is computed within non-overlapping local win-

dows, reducing computational complexity from

quadratic to linear with respect to image size.

2. Shifted Window Partitioning: To ensure con-

nections between windows, a shifted windowing

mechanism is applied in subsequent layers, en-

abling better contextual modeling.

3. Hierarchical Representation: The model cre-

ates multi-scale feature maps by merging neigh-

boring patches as it goes deeper, similar to con-

volutional neural networks (CNNs). This allows

the Swin Transformer to handle tasks requiring

spatial detail, such as semantic segmentation.

1.1.3 Performance and Applications

The hierarchical and eﬃcient design of the Swin

Transformer enables it to excel in various tasks:

Image Classiﬁcation: Achieves a top-1 accu-

racy of 87.3% on ImageNet-1K.

Object Detection: Sets a new benchmark with

a box AP of 58.7 on the COCO dataset.

Semantic Segmentation: Outperforms prior mod-

els with an mIoU of 53.5 on the ADE20K dataset.

These achievements highlight its scalability and ﬂexi-

bility for both basic and dense vision tasks.

1.1.4 Key Contributions

The Swin Transformer brings three major contri-

butions:

Eﬃciency: Linear computational complexity

makes it suitable for high-resolution image pro-

cessing.

Flexibility: Its hierarchical design enables ap-

plications across diverse vision tasks.

Performance: Consistently outperforms state-

of-the-art CNNs and earlier transformer-based

models.

The Swin Transformer represents a transformative

step in computer vision. By addressing the limitations

of earlier Vision Transformers, it sets a new standard

for performance, scalability, and versatility.

1.2 DETR Transformer

The DETR (DEtection TRansformer) is a ground-

breaking object detection framework that simpliﬁes

traditional pipelines by adopting a transformer-based

architecture. Unlike conventional methods that rely on

hand-designed components such as anchor boxes and

non-maximum suppression, DETR treats object detec-

tion as a set prediction problem. This unique approach

ensures that each object is identiﬁed precisely in a sin-

gle, end-to-end process.

1.2.1 Key Features of DETR

Transformer Architecture: DETR employs a

transformer encoder-decoder system, originally

designed for natural language processing, to un-

derstand relationships between objects and their

surrounding context.

1.3 Conclusion

Set Prediction with Bipartite Matching: A bi-

partite matching loss enforces one-to-one corre-

spondence between predicted and ground-truth

objects, eliminating duplicate predictions and

simplifying training.

Simpliﬁed Design: By removing traditional com-

ponents like spatial anchors and post-processing

steps, DETR is conceptually straightforward and

easy to implement with standard tools.

1.2.2 How DETR Works

1. A convolutional neural network (CNN) extracts

features from the input image.

2. The features are ﬂattened and passed to the trans-

former encoder, which captures global image

context.

3. The transformer decoder uses learned object queries

to predict bounding boxes and class labels for ob-

jects in the image.

4. All predictions are generated in parallel, making

the process computationally eﬃcient.

1.3 Conclusion

The advancements in transformer-based architec-

tures like the Swin Transformer and DETR are redeﬁn-

ing the landscape of computer vision by addressing

key challenges and enhancing eﬃciency, scalability,

and versatility. The Swin Transformer excels with its

hierarchical structure and local attention mechanisms,

making it ideal for tasks requiring multi-scale feature

representation. DETR, on the other hand, simpliﬁes

object detection with its end-to-end design and bipar-

tite matching, removing the need for traditional, hand-

crafted components. Together, these models showcase

the immense potential of transformers to unify and ele-

vate vision tasks, paving the way for future innovations

in artiﬁcial intelligence.

References

1. What Are Large Vision Models and How Do

They Work?

2. Swin Transformer: Hierarchical Vision Trans-

former using Shifted Windows

3. End-to-End Object Detection with Transformers

About the Author

Dr. Blesson George presently serves as

an Assistant Professor of Physics at CMS College Kot-

tayam, Kerala. His research pursuits encompass the

development of machine learning algorithms, along

with the utilization of machine learning techniques

across diverse domains.

Domain Generation Algorithm

by Jinsu Ann Mathew

airis4D, Vol.2, No.12, 2024

www.airis4d.com

Imagine trying to catch a criminal who constantly

changes their disguise—this is the challenge cyberse-

curity experts face when dealing with Domain Gener-

ation Algorithms (DGAs). DGAs are algorithms used

by malware to generate a continuous stream of domain

names, which are used to communicate with command-

and-control (C2) servers. These domain names change

frequently, making it nearly impossible for defenders

to block all potential connections. Every time one

domain is blocked, the malware simply switches to

another, staying one step ahead of detection systems.

At their core, DGAs are designed to generate dy-

namic and hard-to-predict domain names using a com-

bination of speciﬁc components. These include the

initial seed value that starts the algorithm, the logic

used to create domain names, and the randomness or

variability that makes them diﬃcult to anticipate. Each

component plays a crucial role in making DGAs eﬀec-

tive tools for evasion. This article focuses on break-

ing down the key components of DGAs, showing how

they work together to support this sophisticated tech-

nique. Understanding these building blocks is essential

to tackling the threat posed by DGA-driven malware.

2.1 Seed Value

A seed value is the initial input to a Domain Gen-

eration Algorithm (DGA) that acts as the starting point

for generating domain names. It’s the core variable that

ensures the algorithm produces a predictable sequence

of domains. Think of it as the ”key” to unlocking the

DGA’s logic. The seed value is designed to be known

to both the malware and its command-and-control (C2)

server so they can independently generate the same list

of domains without any direct communication. The

seed can take many forms, such as:

A hardcoded string embedded in the malware’s

code.

A variable, such as the current date, time, or sys-

tem information.

External data inputs, like stock prices, weather

reports, or even news headlines.

The seed value is critical because it ensures syn-

chronization between the malware and its C2 server

while introducing variability into the domains gener-

ated.

Characteristics of Seed Value

Deterministic : The seed value ensures that the

same input to the algorithm will always produce the

same output. This predictability is vital for synchro-

nization between the malware and its C2 server. For

instance, if the seed is set to ”20241129,” the domains

generated from that seed will always be identical for

both the malware and the server.

Dynamic or Static: Seeds can be Static, where

the value remains ﬁxed. For example, a hardcoded

string like ”malware123” might be used to generate

consistent domains. Alternatively, seeds can be dy-

namic, where the value changes over time or depends

on external factors, such as the current date or system-

speciﬁc data. Dynamic seeds increase the variability

and complexity of the DGA.

Flexible: The seed can be derived from a variety

of sources, including timestamps, random numbers, or

system-speciﬁc parameters like MAC addresses. This

2.2 Domain Generation Logic

ﬂexibility allows attackers to tailor the seed to their

operational needs while ensuring the domains remain

unpredictable to defenders.

2.2 Domain Generation Logic

The domain generation logic is the core mecha-

nism within a Domain Generation Algorithm (DGA)

that transforms inputs, such as a seed value, into a

sequence of domain names. It is essentially the algo-

rithm’s ”blueprint” for creating domain names. This

component dictates how predictable or unpredictable

the generated domains are, playing a crucial role in

evading detection by cybersecurity systems.

The logic is a set of mathematical or program-

matic rules that determine the structure and content

of the generated domains. It processes the input seed

value, combines it with additional variables (like time

or system-speciﬁc data), and produces domains that

adhere to the rules deﬁned in the algorithm.

For example:

1. A seed value (e.g., ”20241129”) is fed into the

logic.

2. The logic applies operations like hashing, con-

catenation, or string manipulations to the seed.

3. The result is a series of domain names, such as

xyz20241129.com or secure 20241129.net.

This logic ensures that both the malware and its

operator’s C2 server can independently generate the

same domains without any direct communication. The

domain generation logic can take various forms to

enhance the unpredictability and resilience of DGA-

generated domains. Simple Concatenation combines

a seed with static strings or numbers, producing do-

mains like malware1.com, malware2.com, which can

be easily anticipated. However, Hash-Based logic ap-

plies a cryptographic or non-cryptographic hash func-

tion to the seed, resulting in seemingly random do-

mains such as x9b2a5.com, making it much harder to

reverse-engineer.

Meanwhile, Time-Based Logic incorporates time

stamps, generating domains like abc20241129.com that

change daily, further complicating blocking eﬀorts.

Lastly, Data-Driven Logic pulls from external factors

like stock prices or weather patterns, generating do-

mains like rainyday2024.net, tied to real-time events.

These diverse techniques ensure that DGA-generated

domains remain highly dynamic and diﬃcult to pre-

dict, making it a challenge for defenders to block them

eﬀectively.

2.3 Entropy Source

An entropy source refers to the input or data used

to introduce randomness into the domain generation

process. The goal of an entropy source is to ensure that

the domains generated are unpredictable and diﬃcult

to anticipate, making it harder for defenders to block

or preemptively take down malicious domains. The

higher the entropy, the greater the unpredictability and

complexity of the generated domains.

An entropy source could be any piece of data

that adds variability to the domain generation process.

Common entropy sources in DGAs include:

Random Numbers: Using a source of random-

ness, such as a pseudorandom number generator (PRNG),

can inject unpredictability into the algorithm. For ex-

ample, random integers might be added to the seed

value to modify the generated domains in an unpre-

dictable manner.

Timestamps: Time-related data, like the current

date or system clock, can serve as a source of entropy.

By incorporating time into the domain generation pro-

cess, the DGA can create domains that change regu-

larly, like daily, making it diﬃcult to predict the next

set of domains.

System Data: Some DGAs use system-speciﬁc

information such as MAC addresses, process identi-

ﬁers (PIDs), or other hardware or software attributes

as entropy sources. This ensures that the generated

domains diﬀer even if the algorithm itself is similar

across diﬀerent systems.

External Data: DGAs can also tap into external

data sources, such as weather reports, stock prices,

or other real-time feeds, to generate domains that are

highly dynamic and tied to real-world events.

The role of the entropy source is crucial for en-

suring that the DGA can generate a large number of

2.4 Frequency and Volume of Domain Generation:

domains, all of which appear random and distinct.

Without a good entropy source, the DGA could pro-

duce easily predictable domains, allowing defenders

to block them more eﬀectively. A high-entropy in-

put ensures that attackers can generate domains that

are nearly impossible for defenders to predict or block

without extensive monitoring.

2.4 Frequency and Volume of

Domain Generation:

Frequency and volume refer to how often and how

many domain names are generated by Domain Genera-

tion Algorithms (DGAs), and both are critical compo-

nents of a DGA’s eﬀectiveness in evading detection and

ensuring continuous communication with a command-

and-control (C2) server.

Frequency refers to how often a DGA generates

new domain names. It can be inﬂuenced by several

factors, including the need for the malware to main-

tain communication with the C2 server. By generating

new domains at regular intervals, the malware can en-

sure that its C2 server remains reachable even if some

domains are blocked by defenders.

Time-based Generation: A DGA may gener-

ate new domains based on time intervals (e.g., daily,

hourly, or even more frequently). For instance, a DGA

might produce a new set of domain names each day,

such as malware20241129.com on one day and mal-

ware20241130.com on the next.

Event-driven Generation: Some DGAs gen-

erate domains when speciﬁc events or triggers occur.

For example, domains might be generated when the

malware detects an attempted block or if it needs to

reconnect to the C2 server after a failed attempt.

Volume refers to the sheer number of domain

names a DGA can generate in a given period. The

volume of generated domains can vary widely, but it is

typically designed to overwhelm detection and block-

ing eﬀorts. A higher volume of domains makes it much

harder for defenders to block all potential communica-

tion channels.

Large Domain Pools: DGAs can generate vast

numbers of domains, creating tens of thousands or

more possible domain names. This increases the chances

that some domains will remain unblocked, allowing the

malware to keep communicating with the C2 server

even if many domains are taken down.

Distributed Generation: Some DGAs can gen-

erate multiple domains simultaneously, creating di-

verse, redundant communication paths for the malware

to use. This distributed approach ensures that, even if

certain domains are blocked, other domains will be

available for the malware to contact the C2 server.

The combination of high volume and frequent do-

main generation complicates eﬀorts to block or disrupt

communication. Even if defenders identify and block

some of the domains, the DGA can quickly generate

new ones to bypass these defenses.

2.5 Conclusion

In conclusion, Domain Generation Algorithms

(DGAs) are a common technique used by malware

to stay hidden and maintain communication with at-

tackers. By generating large numbers of new domain

names at regular intervals, DGAs make it hard for de-

fenders to block all potential malicious domains. The

key components of DGAs, such as seed values, domain

generation logic, and entropy sources, work together to

create unpredictable and ever-changing domains. The

frequency and volume at which these domains are gen-

erated further complicate eﬀorts to block them. Un-

derstanding how these components function helps in

developing better defense strategies to protect against

this evolving cyber threat.

References

Domain generation algorithm

DGA Detection with data analytics

What is Domain Generation Algorithm?

Real-Time Detection of Dictionary DGA Net-

work Traﬃc Using Deep Learning

2.5 Conclusion

About the Author

Jinsu Ann Mathew is a research scholar

in Natural Language Processing and Chemical Infor-

matics. Her interests include applying basic scientiﬁc

research on computational linguistics, practical appli-

cations of human language technology, and interdis-

ciplinary work in computational physics.

How to Improve Your ML models

by Linn Abraham

airis4D, Vol.2, No.12, 2024

www.airis4d.com

3.1 Introduction

The primary goal of developing a machine learn-

ing model is to be able to learn from training sam-

ples and generalize this learning to real world samples.

When developing machine learning models, you are

often in a better situation if your model is overﬁtting

rather than if its underﬁtting. That is why achieving

overﬁtting is in itself considered an achievement or

rather as a ﬁrst step towards making progress. When

we talk about generalization we talk about a model that

is overﬁt. How do we improve the performance of an

overﬁtting model or rather how do we improve the gen-

eralization of our network? But before answering that,

let us ask ourselves why Machine Learning or Deep

Learning works at all? Is Machine Learning just glori-

ﬁed interpolation or is there even any relation between

the two? Understanding the answer to these would also

allow us to appreciate how human intelligence is seen

to better than this so called machine intelligence. We

will also come to certain practical considerations for

improving the generalization of our models.

3.2 Generalization in ML models

Consider the input space in a problem as simple

as the MNIST classiﬁcation problem. The images here

are 28 × 28 arrays of integers between 0 and 255. The

total number of elements in the input space, which

is 256 to the power of 784, is much greater than the

number of atoms in the universe. Whereas the number

of training samples in the MNIST data is around 60

thousand. So how is it possible to achieve near perfect

Figure 1: Top panel shows how a model goes from

underﬁt to overﬁt. The bottom panel shows how the two

diﬀerent models perform on test data. Image Credit:

Francois Chollet.

accuracies on this problem with networks of moderate

size? Is interpolation or memorization going to help

in such a case? However not every point in the input

space would be a realistic input, i.e. to say, resemble a

handwritten digit. The subspace of such valid MNIST

samples infact occupy a tiny region of the original space

and one which is highly structured.

Let us try to understand this structure. First, this

subspace is a continous subspace meaning that if you

take any sample and modify it in a tiny little way, it will

still be recognizable as the same handwritten digit.

Second, there is connectedness, if you take any two

random MNIST digits A and B, there exists a sequence

of “intermediate” images that morph A into B in such

a way that the consecutive images resemble the same

digit. This is not to say that there can’t be ambiguous

shapes close to the boundary between the two digits.

The technical term for such a subspace is a “man-

ifold”. Just like how a smooth curve in the plane is

said to be a 1D manifold within a 2D space. Because

3.3 How to Improve Generalization

Figure 2: A sample of handwritten digits in the

MNIST dataset. Image Credit: Wikipedia

for every point on the curve, you can draw a tangent

or in other words, the curve can be approximated by a

line at every point. Similiar reasons motivate you to

call a smooth surface as a 2D manifold within a 3D

space. Thus to summarize the discussions so far, the

manifold hypothesis posits that all natural data lies on a

low-dimensional manifold within a high-dimensional

space. This is not just true for MNIST digits, but also

for human faces, the structure of galaxies, the sounds

of musical instruments, and even molecular conﬁgura-

tions.

The consequences of this manifold hypothesis is

that:

Machine learning models work in this low-dimensional

highly structured subspaces within the actual in-

put space of the problem, which can be called

a latent manifold. This is also the kind of thing

that happens when cleverly designed features are

designed for seemingly diﬃcult problem.

Training data and real world data all consists of

a random sampling of points from this manifold.

The model then tries to interpolate between two

inputs. That is, morph one into another via a

continous path within the manifold.

This interpolation ability of machine learning mod-

els seems to be just one of the many ways to achieve

generalization. Because humans, as we know, are ca-

pable of extreme generalization when it comes to han-

dling everyday situations. This is without the need to

be pre-trained on these situations. This comes from

our ability to reason, making use of cognitive mech-

anism such as abstraction, symbolic representation of

the world, logic, common sense and innate priors about

the world.

3.3 How to Improve Generalization

We saw that the ability of ML models to generalize

is more of a consequence of the natural structure of the

data than any property of the model itself. This is

why the most bang for buck improvement in model

performance comes from spending time with the data

- cleaning data, using more informative features and

removing noisy features. While adding more data you

need to make sure that this addition is leading to a

more dense sampling of the input space. It is when

all these steps have been exhausted that you should

consider model improvements. With regard to model

and training you can do the following to maximize

generalization or to avoid overﬁtting.

Stop before you overﬁt. This involves using early

stopping and saving your model to use the best

possible model.

Regularize your model. This involves a set of

best practices to prevent your model from over-

ﬁtting. The end goal is to make your model more

smooth or ‘regular’.

3.4 Regularization Techniques

If your model is overparametrized try reducing

the network size. At the same time make sure that you

do not go into underﬁtting. With weight regularization

techniques you prefer a model that relies on a lesser

number of connections. This is done by penalizing the

magnitude of the weights propotional to the absolute

magnitude (l1-regularization) or to the square of the

magnitude (l2-regularization). However note that this

doesn’t have much of an eﬀect if your model is very

heavily parametrized in the ﬁrst place. Hence this

is much more commonly used with smaller models.

For larger models dropout is an eﬀective technique for

regularizations. This involves randomly removing a

diﬀerent set of neurons on each example. This would

impede the ability of the network to learn patterns that

aren’t signiﬁcant. The dropout rate decides the number

REFERENCES

of neurons to be removed.

References

[Chollet(2021)] Franc¸ois Chollet. Deep Learning with

Python. Manning, Shelter Island, NY, second edi-

tion edition, 2021. ISBN 978-1-61729-686-4.

About the Author

Linn Abraham is a researcher in Physics,

specializing in A.I. applications to astronomy. He is

currently involved in the development of CNN based

Computer Vision tools for prediction of solar ﬂares

from images of the Sun, morphological classiﬁca-

tions of galaxies from optical images surveys and ra-

dio galaxy source extraction from radio observations.

Part II

Astronomy and Astrophysics

Black Hole Stories-14

Gravitational Wave Emission

by Ajit Kembhavi

airis4D, Vol.2, No.12, 2024

www.airis4d.com

1.1 Introduction

In the next few stories we will consider gravita-

tional wave emission, the ﬁrst detection of gravitational

waves by the LIGO detectors, the many merging black

hole and neutron star binaries which have been dis-

covered since then as sources of gravitational waves,

and their implications for black hole astrophysics. In

the present story we summarise some elements of elec-

tromagnetism and electromagnetic waves, which will

greatly help in understanding through analogies the

more complex facts about gravitational waves. We will

limit ourselves to only those matters about electromag-

netism which are directly relevant to the subsequent

stories.

A detailed and yet widely accessible discussion

on gravitational waves and electromagnetic radiation

can be found in the book Gravitational Waves: A New

Window to the Universe, by Ajit Kembhavi & Pushpa

Khare, Springer (2020) and a more technical and yet

very lucid discussion is the book Gravity, An Intro-

duction to Einstein’s General Relativity by James B.

Hartle, Pearson Education (2002). A recent preprint

of the book has been published by Cambridge Uni-

versity Press. An excellent introduction to radiation

processes is the book Radiative Processes in Astro-

physics by George B. Rybicki and Alan P. Lightman,

Wiley (1979).

1.2 Electromagnetic Waves

In this section we will very brieﬂy summarise the

generation and properties of electromagnetic waves,

which will help us to understand the more complex

nature of gravitational waves.

Electromagnetic phenomena are the eﬀects of elec-

tric and magnetic ﬁelds. The ﬁelds are produced by

electric charges and electric currents, the latter being

the produce by moving electric charges. The electric

and magnetic ﬁelds exert a force on electric charges.

The ﬁelds produced by a distribution of charges and

currents are described by Maxwell’s equations of elec-

tromagnetism. These are four vector equations, which

were developed by James Clerk Maxwell in 1865. All

electromagnetic phenomena can be derived from these

four equations, and the Lorentz force equation which

provides the force exerted by electric and magnetic

ﬁelds on a moving electric charge.

Maxwell’s ﬁeld equations describe laws which

were known before 1865, like Faraday’s which relates

the current produced in a loop of wire by a magnet

passing through the loop, or the magnetic ﬁeld pro-

duced around a current carrying wire, which is the

Biot-Savart law. But the equations also predict com-

pletely new phenomena, the most important of which

is the existence of electromagnetic waves. From the

equations it can be easily shown that in regions of

space which are free of electric charges and currents,

electric and magnetic ﬁelds satisfy a wave equation and

propagate as waves through the space. The speed of

1.3 Radiation from an Accelerated Charge

propagation is equal to the speed of light. It follows that

light must in fact be an electromagnetic wave. Electro-

magnetic waves were ﬁrst experimentally detected by

Henrick Hertz in the period 1886-89; he was able to

show that the waves indeed travelled with the speed of

light.

To calculate the ﬁelds produced by a given dis-

tribution of charges and currents, it is convenient to

introduce a vector A(x, t), known as the vector poten-

tial. and another scalar function ϕ((x, t). The electric

ﬁeld E and magnetic ﬁeld B can be obtained from A and

ϕ by taking space and time derivatives. The equations

for appropriately chosen potentials through a process

known as gauge transformation are

where the ∇

Φ is given by

∇

Φ =

∂

∂x

∂

∂y

∂

∂z

with similar expressions for the three components

Ax, Ay, Az of the vector A. ρ is the density of electric

charge and j is the current density. The diﬀerential

equations have the familiar form of wave equations in

physics, and can be solved for given charge and current

densities, with appropriate boundary conditions. The

electric and magnetic ﬁelds can then be obtained from

the potentials.

The simplest case of solutions of the wave equa-

tion is a charge q at rest in some inertial frame

reference. There is just the static electric ﬁeld given

by Coulomb’s law. The next simple case is that of

a charge moving with a uniform velocity. Since this

constitutes a current, both electric and magnetic ﬁelds

are present. The ﬁelds at a given point change with

time as the charge moves away from it, or towards it,

but their form is not wave like. To obtain wave solu-

tions which carry energy to large distance we have to

consider ﬁelds produced by an accelerated charge.

An inertial frame is a frame of reference in which a particle at rest

will remain at rest and a particle moving with uniform velocity will

continue to do so. The special theory of relativity requires that all

laws of physics have the same form in any inertial frame.

1.3 Radiation from an Accelerated

Charge

The electric and magnetic ﬁelds of a single charge

moving with acceleration, i.e. with changing velocity,

can be obtained from the wave equations. It turns out

that the electric ﬁeld is a sum of two parts: the ﬁrst

known as a velocity ﬁeld is just the ﬁeld for a charge in

uniform motion, while the second part, known as the

acceleration ﬁeld is proportional to the acceleration of

the particle. The magnetic ﬁeld has the same properties

and is perpendicular to the electric ﬁeld. An important

aspect of these ﬁeld is that they are inversely propor-

tional to the distance from the charge. A consequence

is that the ﬁelds carry energy to great distances from

the charge, and an accelerated charge therefore is said

to radiate energy. In the special case of a charge with

velocity much smaller the velocity of light, the power

radiated by the charge is given by Larmor’s formula:

where P is the energy radiated per second, q is the

electric charge, a the acceleration and c the speed of

light. In the process the charge will lose energy and

its acceleration will reduce, unless the energy of the

charge is replenished in some way.

The radiation is not emitted equally in all direc-

tions; the emission in a direction making an angle θ

with the direction of the acceleration is proportional to

sin

θ, as shown in Figure 1:

Figure 1: The pattern of radiation from an ac-

celerated charge moving with velocity much smaller

than the velocity of light. The acceleration is along

the z-axis. Maximum emission occurs in the direction

perpendicular to the acceleration direction. The three

1.4 Plane Electromagnetic Waves

dimensional pattern is obtained by rotating the ﬁgure

around the z-axis. Figure courtesy Shimon Levit.

When the velocity of the radiating particle ap-

proaches the velocity of light, the expression for the

emitted power is

Which is the special relativistic form for radia-

tion from an accelerated charge. Here a

⊥

and a

∥

are

the components of the acceleration perpendicular and

parallel to the instantaneous velocity of the charge, and

is the Lorentz factor of special relativity. When

the velocity of the charge γ is much smaller than the

velocity of light c, γ→1 and the relativistic form re-

duces to Larmor’s formula. In the relativistic case the

angular distribution depends on the direction between

the vector velocity of the charge and its acceleration

vector. The pattern for a highly relativistic charge with

acceleration perpendicular to the velocity is shown in

Figure 2:

Figure 2: Angular distribution of radiation emit-

ted by a charge moving with velocity approaching the

velocity of light. The maximum radiation occurs in

the direction of motion. The radiation is mainly con-

ﬁned to a cone with opening angle ˜ 1/γ. The cone

becomes narrower as the velocity increases, since γ

increases with velocity. The diagram is from Rybicki

and Lightman (1979).

The radiation is clearly beamed in the forward

direction, with the beam becoming narrower as the

velocity increase. Such is the radiation emitted by

highly energetic accelerated electrons in astrophysical

sources like quasars, and beaming is responsible for

many interesting phenomena observed in such sources.

1.4 Plane Electromagnetic Waves

In free space, where there are no charges or cur-

rents, it can be easily shown from Maxwell’s equations

that the electric ﬁeld E and the magnetic ﬁeld B, which

are three dimensional vectors, satisfy wave equations

of the form of discussed above, with the right hand

side being zero. We will consider as an example a very

simple solution of the form

At a given time t, such a ﬁeld has the form of

a sine wave as a function of the spatial coordinate x.

The distance between successive crests (maxima) or

troughs (minima) in the wave is the wavelength λ =

2π/k. At a given point x, the ﬁelds oscillate in time as

a sine function, with the number of oscillations per unit

time being the frequency ν = ω/2π. E

is the amplitude

of the vibration of the electric ﬁeld and B

of the

magnetic ﬁeld. It follows from the ﬁeld equation that

(1) the magnitudes of the amplitudes are equal, E

, (2) E and B are perpendicular to each other and

to the x-direction in which the wave is propagating and

(3) νλ = c and therefore the wave propagates forward

with speed c. This is known as a plane wave solution

because the vibrations of the ﬁelds are conﬁned to a

plane, in this case the yz plane. More generally we

can consider a plane wave propagating with speed c in

some direction deﬁned by a unit vector n. Again E and

B are perpendicular to each other and to n.

The electromagnetic ﬁeld carries energy and mo-

mentum, and in vacuum has the energy density U =

+ B

)/8π , where E and B are magnitudes of

the vector ﬁelds. The ﬂux of electromagnetic energy

across a surface at a given point in space is the vector

cross product S = c(E x B)/4π. For plane waves, these

reduce to U = E

/4π and S = cU, the last expression

showing that the energy ﬂows with speed c. The above

results are all valid for waves propagating in vacuum.

When the waves propagate though a material medium,

properties which describe the medium enter various

expressions.

The wave equations being linear, the sum any so-

lutions of the equations is also a solution. We can

therefore consider the simultaneous propagation in a

1.5 Polarisation

given direction of waves with diﬀerent frequencies (we

can also more generally consider waves travelling in

diﬀerent directions). Depending on how the peaks

and troughs of the waves being added are situated, the

waves can interfere constructively in some places and

destructively in others. The result is that instead of

having an inﬁnite train of sinusoidal waves, we get

more localised wave packets which propagate forward

with velocity c. Because now we have a packet, this is

known as the group velocity.

Astronomical source often emit radiation in the

form of pulses, which travel through space as a wave

packet. It is possible to carry out Fourier analysis of

such a packet to ﬁnd the component frequencies of

the plane waves which make up the packet. Such an

analysis is necessary since diﬀerent frequencies can be

produced by diﬀerent physical processes in a source,

and ﬁnding the components provides insight into the

radiation emitting mechanisms and their distribution in

the source.

In Figure 3 is shown a wave packet, in which

the electric ﬁeld E (and therefore magnetic ﬁeld B)

is a sine wave of frequency ν

= ω

/2π over a ﬁnite

time interval T. Outside of T, the ﬁeld is zero, which

shows that other frequency components besides ν

are

present. These interfere destructively in such a way

that the ﬁeld is non-zero only over the time interval T,

producing the travelling wave packet. Fourier analysis

of the packet produces the distribution of frequencies

shown in Figure 4, with the power corresponding to

each frequency shown along the y-axis. The maximum

contribution comes from ω

, but sine waves of a whole

range of frequencies are present, distributed on either

side of the main frequency.

Figure 3: A sine wave packet of ﬁnite duration

T. See text for details. The ﬁgure is from Rybicki &

Lightman (1979)

Figure 4: The frequency distribution of sine wave

components of the wave packet in Figure 2 . See text

for details. The ﬁgure is from Rybicki & Lightman

(1979)

The range of frequencies in the electromagnetic

spectrum is vast. At the lowest frequencies, i. e., at

the longest wavelengths there are radio waves; as the

frequency increases, we pass through diﬀerent regions

of the spectrum which are shown in Figure 4. Astro-

nomical source like quasars emit copiously at all these

wavelengths, which indicates that a range of emitting

mechanism is present in them. There is a factor of ˜10

in the range of frequencies observed in such sources.

Figure 5: Frequency and wavelength for diﬀerent

regions of the spectrum, ranging from <10

Hz in the

radio region to >10

Hz for high energy Gamma rays.

Figure from https://www.miniphysics.com/electromagnetic-

waves.html.

1.5 Polarisation

In the case of a plane wave travelling in the x-

direction, we have seen that the electric and magnetic

ﬁelds oscillate in a ﬁxed direction. The two ﬁelds are

always perpendicular to the direction of propagation

and to each other. We can chose the x- and y-axes such

that E is along the x-axis and B is along the y-axis.

Since the two ﬁelds are perpendicular to each other,

we can simply consider the E ﬁeld in the following

1.5 Polarisation

discussion, with similar remarks applying to the B ﬁeld.

Since the electric ﬁeld oscillates in a ﬁxed direc-

tion, it is said to be linearly polarised. The direction of

oscillation and of propagation together deﬁne a plane

which is known as the plane of polarisation. Now sup-

pose another plane wave is present, with the electric

ﬁeld now being in the y-direction. The sum of the two

waves is also a plane wave solution, with the behaviour

of the total electric ﬁeld depending on the magnitude

and phase relation between the two ﬁelds. If the elec-

tric ﬁelds have the same magnitude and pass through

maxima and minima at the same time, they are said to

be in phase. Then the magnitude of the resultant ﬁeld

is given by E = (E

+ E

)

1/2

and the ﬁeld is linearly

polarised at an angle of 45 degrees to the x- and y-axes.

If there is a phase diﬀerence of 90 degrees, i.e. when

one wave passes through a maximum the other passes

through a zero, the magnitude of the resultant ﬁeld is

again given by E = (E

)

1/2

, but the tip of the vector

now traces a circle, rotating in the clockwise direction,

as seen by an observer towards whom the wave propa-

gates. For a phase diﬀerence of -90 degrees, the wave

vector traverses a circle, rotating in the anti-clockwise

direction. In these two cases the wave is said to be cir-

cularly polarised. For a phase diﬀerence intermediate

between the two values we have considered, the tip of

the electric vector traces a clockwise or anti-clockwise

ellipse, and the wave is elliptically polarised. Circular

and linear polarisation can be considered to be special

cases of elliptical polarisation over the whole range

from -90 degrees to +90 degrees.

: In the next story we will consider gravitational

waves, which because of the nature of Einstein’s equa-

tions of general relativity are so much harder to study.

The concepts developed for electromagnetic waves will

make that process somewhat easier.

About the Author

Professor Ajit Kembhavi is an emeritus

Professor at Inter University Centre for Astronomy

and Astrophysics and is also the Principal Investiga-

tor of the Pune Knowledge Cluster. He was the former

director of Inter University Centre for Astronomy and

Astrophysics (IUCAA), Pune, and the International

Astronomical Union vice president. In collaboration

with IUCAA, he pioneered astronomy outreach activ-

ities from the late 80s to promote astronomy research

in Indian universities.

Unveiling the Stars: A Deep Dive into Stellar

Spectral Classiﬁcation

by Sindhu G

airis4D, Vol.2, No.12, 2024

www.airis4d.com

2.1 Introduction

Stars, the building blocks of galaxies, are more

than mere points of light; they are complex cosmic lab-

oratories that reveal the fundamental processes of the

universe. Stellar spectral classiﬁcation is a systematic

method used by astronomers to categorize stars based

on their light, primarily their spectra. By analyzing this

light, split into a spectrum of colors interspersed with

spectral lines, scientists can determine a star’s temper-

ature, composition, luminosity, and evolutionary stage.

This classiﬁcation, which summarizes the ioniza-

tion state of a star’s atmosphere, helps decode the phys-

ical and chemical properties of stars and their roles in

the structure of galaxies. Through methods such as

using prisms or diﬀraction gratings, astronomers iden-

tify spectral lines corresponding to speciﬁc elements or

molecules, with their strengths varying primarily due

to the temperature of the star’s photosphere.

Stellar spectral classiﬁcation is a foundational tool

in astronomy, oﬀering insights into the life cycles of

stars and enabling a deeper understanding of the uni-

verse. This article explores the history, methodology,

specialized classes, and modern applications of this

essential system.

2.2 Historical Development of Stellar

Spectral Classiﬁcation

The classiﬁcation of stars began in the 19th cen-

tury when astronomers studied stellar spectra to un-

cover their properties. In 1814, Joseph von Fraunhofer

identiﬁed dark absorption lines in the solar spectrum,

revealing the chemical composition of the Sun. By

the late 19th century, Angelo Secchi divided stars into

four spectral classes based on visual diﬀerences in their

spectra, laying the groundwork for systematic classiﬁ-

cation.

The modern classiﬁcation system emerged at Har-

vard College Observatory under Edward C. Pickering

and his team, including Annie Jump Cannon. Cannon

simpliﬁed earlier schemes into the Harvard Classiﬁ-

cation System, which organizes stars by surface tem-

perature in a sequence from O to M, representing the

hottest to the coolest stars. Cecilia Payne-Gaposchkin

further advanced the ﬁeld by linking spectral lines to

stellar composition, forming the basis for modern as-

trophysics.

By the 1940s, the Morgan-Keenan (MK) system

built on this foundation, integrating temperature and lu-

minosity into a comprehensive framework that remains

the standard for stellar classiﬁcation today.

2.3 The Basics of Stellar Spectral

Classiﬁcation

The classiﬁcation of stars is primarily based on

their spectra—light spread into its component colors

or wavelengths, much like a rainbow. When starlight

passes through a prism or diﬀraction grating, dark ab-

sorption lines appear, corresponding to speciﬁc ele-

2.4 The Harvard Spectral Classiﬁcation System

ments in the star’s atmosphere. The patterns and

strengths of these lines are key to determining a star’s

spectral type.

2.4 The Harvard Spectral

Classiﬁcation System

The Harvard Classiﬁcation System, developed in

the early 20th century, organizes stars by their surface

temperature and spectral features. Stars are grouped

into seven main types—O, B, A, F, G, K, and M—ranging

from the hottest and most massive (O-type) to the

coolest and least massive (M-type). The spectral classes

O through M, along with other specialized categories,

are divided into subcategories numbered 0 to 9, with

0 representing the hottest stars within a class and 9 the

coolest. For instance, A0 represents the hottest stars in

the A class, while A9 indicates the coolest. Fractional

values are also used, such as the classiﬁcation of the

star Mu Normae as O9.7. The Sun is classiﬁed as G2.

O-type stars are extremely hot and blue, with tem-

peratures exceeding 30,000 K and spectra dominated

by ionized helium lines. B-type stars are slightly cooler

(10,000–30,000 K) and show neutral helium lines. A-

type stars (7,500–10,000 K) are white, characterized

by strong hydrogen absorption lines, while F-type stars

(6,000–7,500 K) are yellow-white with weaker hydro-

gen lines and prominent ionized metal lines. G-type

stars, like the Sun (5,200–6,000 K), are yellow with

spectra dominated by ionized calcium and metals. K-

type stars (3,700–5,200 K) are orange with strong neu-

tral metal lines, and M-type stars, cooler than 3,700 K,

are red with molecular bands like titanium oxide.

Cecilia Payne-Gaposchkin demonstrated that the

sequence of O-B-A-F-G-K-M reﬂects a temperature

progression. Her work, building on Meghnad Saha’s

ionization theory, showed that the spectral lines indi-

cate both the temperature of a star’s photosphere and

its chemical composition. While the system’s subdivi-

sion into subtypes is based on absorption line strengths,

these divisions are not evenly spaced. Mnemonics such

as ”Oh Be A Fine Girl/Guy, Kiss Me” help remember

the sequence. The Harvard system remains fundamen-

Figure 1: Harvard Spectral Classiﬁcation. (Image

Credit: Ram Chandra Gotame and Physicsfeed.com)

tal to understanding stellar properties and their role in

the universe.

2.5 Yerkes Spectral Classiﬁcation

The Yerkes, or MK (Morgan-Keenan), classiﬁ-

cation expands on the Harvard system by incorporat-

ing stellar size and brightness. Introduced in 1943 by

William Morgan, Philip Keenan, and Edith Kellman,

it is a two-dimensional system based on spectral lines

sensitive to surface temperature and gravity, which cor-

relates with luminosity. While the Harvard system fo-

cuses on temperature, the MK system adds Roman nu-

merals to denote luminosity classes, categorizing stars

as follows:

I: Supergiants (e.g., Betelgeuse)

Ia-0: hypergiants or extremely luminous

supergiants

Ia: luminous supergiants

Iab: supergiants

Ib: less luminous supergiants

II: Bright giants

III: Giants (e.g., Aldebaran)

IV: Subgiants

V: Main-sequence stars (e.g., the Sun)

Va: extremely luminous dwarfs

Vab: luminous dwarfs

Vb: normal dwarfs

Vz: lower main sequence dwarfs

VI: Subdwarfs

VII: White dwarfs

2.6 Specialized Stellar Spectral Classes

For example, the Sun is classiﬁed as G2V, indi-

cating it is a G-type main-sequence star with a sur-

face temperature of about 5,800 K. Denser stars with

higher surface gravity exhibit broader spectral lines

due to pressure, while giant stars with lower gravity

show narrower lines, allowing luminosity eﬀects to be

deduced from the spectrum.

Marginal cases are marked using additional sym-

bols: a slash (/) for stars belonging to either of two

classes (e.g., III/IV) and a dash (-) for stars in transi-

tional stages (e.g., A3-4III/IV). Subclasses such as IIIa

or IIIb indicate slight variations in brightness within a

class. Though nominal luminosity class VII was once

used for white dwarfs, it is now rare, as white dwarfs

are classiﬁed separately from the temperature-based

sequences used for main-sequence and giant stars.

2.6 Specialized Stellar Spectral

Classes

Not all stars ﬁt into the standard O-M classiﬁca-

tion due to their unique properties, leading to special-

ized categories:

Wolf-Rayet (WR) Stars: Hot, massive stars

with broad emission lines caused by intense stel-

lar winds, showing helium, carbon, and oxygen

but lacking hydrogen.

Carbon Stars (C-type): Cool red giants with

carbon-rich atmospheres, characterized by strong

molecular bands of carbon compounds.

L, T, and Y Dwarfs: Substellar objects, includ-

ing brown dwarfs, cooler than M dwarfs, with

spectra dominated by methane and water vapor.

Peculiar Stars (e.g., Ap, Bp): Stars with un-

usual chemical abundances or strong magnetic

ﬁelds that alter their spectra.

Zirconium Giant Stars (S-type): S-type stars

are a class of cool giants that are characterized by

an abundance of s-process elements, especially

zirconium. Class S stars form a continuum be-

tween class M stars and carbon stars. They are

similar to M-type giants, but they have a higher

proportion of these heavy elements in their spec-

tra.

These specialized classes capture the diversity of

stellar phenomena and evolutionary paths beyond the

O-M sequence.

2.7 The Physics of Stellar Spectra

Stellar spectra form as light interacts with a star’s

outer layers, where atoms and molecules absorb spe-

ciﬁc wavelengths, producing absorption lines linked to

atomic transitions. These lines reveal crucial proper-

ties:

Surface Temperature: Deduced from the peak

wavelength, color, and strengths of spectral lines.

Chemical Composition: Identiﬁed by charac-

teristic absorption lines of elements like hydro-

gen, helium, calcium, and iron.

Radial Velocity: Measured through Doppler

shifts in spectral lines, indicating motion toward

or away from Earth.

Magnetic Fields: Detected via the Zeeman ef-

fect, which causes spectral line splitting.

The strength and presence of these lines depend

on temperature, pressure, and ionization levels, pro-

viding insights into a star’s physical conditions, struc-

ture, and evolution. Excitation balance describes how

atoms absorb energy and shift between energy levels,

while ionization balance reﬂects the degree of element

ionization in a star’s atmosphere, which rises with in-

creasing temperature. The varying strengths of these

spectral lines enable astronomers to determine a star’s

temperature, chemical composition, and surface grav-

ity.

2.8 Applications of Spectral

Classiﬁcation

Stellar spectral classiﬁcation plays a crucial role

in various areas of astronomy:

2.9 Advances in Spectral Classiﬁcation

Stellar Evolution: By combining spectral type

and luminosity, astronomers trace a star’s life

cycle, from its formation as a protostar to its ﬁnal

state as a white dwarf, neutron star, or black hole.

Galactic Structure: Large surveys like the Sloan

Digital Sky Survey (SDSS) and Gaia map stellar

distributions to study the Milky Way’s structure

and dynamics.

Exoplanet Research: Classifying host stars is

essential for understanding exoplanet environ-

ments, especially in the study of habitable zones,

with M dwarfs being common targets.

Discovery of Rare Objects: Automated classi-

ﬁcation systems have uncovered rare stellar ob-

jects such as hypervelocity stars and extremely

metal-poor stars.

Additionally, spectral classiﬁcation aids in Pop-

ulation Studies by analyzing star populations within

galaxies and helps measure stellar Distance through

luminosity, enabling distance estimates using methods

like the inverse square law.

2.9 Advances in Spectral

Classiﬁcation

Modern astronomy has revolutionized spectral clas-

siﬁcation through technological advancements:

Automated Surveys

Instruments like Gaia and ASAS-SN capture mil-

lions of stellar spectra, enabling large-scale clas-

siﬁcation.

Machine Learning

Algorithms process spectral data to classify stars,

detect anomalies, and identify rare phenomena.

High-Resolution Spectroscopy

Advanced spectrographs, such as those on the

James Webb Space Telescope (JWST), reveal

unprecedented details of stellar spectra.

Integration with Astroseismology

Combining spectral data with stellar oscillation

studies provides precise stellar parameters.

These innovations expand our ability to study stars

and the broader universe.

2.10 Conclusion

Stellar spectral classiﬁcation is a powerful tool

that transforms light into knowledge. From the Harvard

and Yerkes systems to modern automated techniques,

it remains a cornerstone of astrophysical research. By

categorizing stars, astronomers unravel their physical

properties, life cycles, and contributions to the uni-

verse’s evolution. As new technologies and method-

ologies emerge, the classiﬁcation of stars will continue

to illuminate the mysteries of the cosmos, ensuring our

understanding of the universe grows ever deeper.

References:

Spectral Classiﬁcation of Stars

Stellar classiﬁcation

Spectral Classiﬁcation of Stars

Morgan-Keenan Luminosity Class

Spectral Classes

A note on the spectral atlas and spectral classiﬁ-

cation

Stars and their Spectra

The classiﬁcation of stellar spectra

The Classiﬁcation of Stellar Spectra

About the Author

Sindhu G is a research scholar in Physics

doing research in Astronomy & Astrophysics. Her

research mainly focuses on classiﬁcation of variable

stars using diﬀerent machine learning algorithms. She

is also doing the period prediction of diﬀerent types

of variable stars, especially eclipsing binaries and on

the study of optical counterparts of X-ray binaries.

Part III

Biosciences

Innovating Autism Support: AI-Driven

Approaches for a More Inclusive Society

by Kalyani Bagri

airis4D, Vol.2, No.12, 2024

www.airis4d.com

1.1 Introduction

Autism Spectrum Disorders (ASD) represent a

diverse group of neurodevelopmental conditions char-

acterized by challenges in social interaction, commu-

nication, and repetitive behaviors. The spectrum na-

ture of ASD reﬂects varying degrees of severity, im-

pacting individuals diﬀerently and presenting unique

developmental trajectories. Typically manifesting in

early childhood, ASD often coexists with conditions

like epilepsy, anxiety, depression, or attention deﬁcit

hyperactivity disorder (ADHD).

Historically, ASD was misunderstood and often

misdiagnosed, sometimes being confused with schizophre-

nia. This misclassiﬁcation led to ineﬀective interven-

tions and stigma. However, signiﬁcant advances in ge-

netics and neuroscience during the 20th century have

reshaped our understanding, moving from psychoso-

cial explanations to a biomedical model that empha-

sizes genetic and neurobiological underpinnings.

Despite these strides, challenges persist especially

in countries like India, where access to specialized care

is limited. Growing awareness and the integration of

advanced technologies, such as Artiﬁcial Intelligence

(AI), oﬀer opportunities to improve ASD diagnosis,

treatment, and management.

1.2 Understanding the Causes of ASD

To better address ASD, it is crucial to comprehend

its multifaceted causes. ASD results from a complex

interplay of genetic, environmental, and neurobiologi-

cal factors. These factors include:

Genetic Factors: Genetic variations, ranging

from single-gene mutations to complex multi-

gene interactions, play a critical role in the de-

velopment of ASD. Speciﬁc genetic syndromes,

such as Fragile X and Rett syndrome are associ-

ated with ASD.

Environmental Factors: Prenatal exposure to

certain environmental risks, including maternal

infections, medications, and pollutants, may in-

crease the likelihood of ASD. Nutritional de-

ﬁciencies and advanced parental age have also

been implicated.

Neurobiological Factors: Structural abnormal-

ities in the brain, neurotransmitter imbalances,

and atypical neural connectivity patterns con-

tribute to ASD. These neurobiological disrup-

tions aﬀect cognitive and emotional functioning.

Understanding these causes is vital in developing

eﬀective interventions. Consequently, early diagnosis

and intervention play a critical role in mitigating the

impact of ASD.

1.3 Importance of Early Diagnosis and Intervention

1.3 Importance of Early Diagnosis

and Intervention

Building on this understanding, early diagnosis

and intervention are pivotal in providing access to tai-

lored interventions that optimize developmental out-

comes. Early diagnosis is vital in mitigating the im-

pact of ASD, reducing symptoms, and lessening emo-

tional and ﬁnancial burdens on families. Furthermore,

delayed identiﬁcation can exacerbate symptoms and

heighten emotional and ﬁnancial burdens on families.

Consequences of Delayed Diagnosis

Delayed Intervention: Reduced opportunities for

early therapies, such as Applied Behavior Anal-

ysis (ABA).

Increased Severity of Symptoms: Behavioral

and cognitive challenges intensify over time.

Impact on Families and Caregivers: Emotional

stress and ﬁnancial strain can aﬀect caregiver’s

well-being.

Beneﬁts of Early Intervention Programs

In contrast, early intervention programs can sig-

niﬁcantly improve communication, emotional regula-

tion, and adaptive functioning. Such programs include:

Speech therapy

Occupational therapy

Social skills training

Moreover, AI’s potential to revolutionize early

diagnosis and intervention extends beyond individual

beneﬁts, oﬀering broader societal advantages. These

advantages include reducing economic and caregiv-

ing burdens while promoting inclusivity and empow-

erment for individuals with ASD.

1.4 Transforming Autism Diagnosis

and Care with AI

In recent years, AI has revolutionized the diagno-

sis, treatment, and management of ASD. AI-powered

tools oﬀer innovative solutions that address traditional

challenges, such as delayed detection and subjective

assessments.

Revolutionizing Diagnosis

The application of AI in diagnosis has been particularly

noteworthy. Natural Language Processing (NLP) tech-

niques detect communication deﬁcits and social cues,

enabling earlier and more accurate diagnosis. This, in

turn, is critical for eﬀective intervention and treatment.

Personalized Therapies

In addition to diagnosis, AI-powered therapy platforms

have transformed the treatment landscape. These plat-

forms tailor personalized plans to individual needs,

continuously adapting and adjusting based on progress

and response. Virtual therapists and simulated environ-

ments play a crucial role in social skills development.

Enhancing Ongoing Management

Furthermore, AI plays a vital role in enhancing ongo-

ing management of ASD. Predictive analytics antici-

pate potential challenges, empowering caregivers to act

preemptively. Mobile apps equipped with AI provide

caregivers with valuable resources to support ongoing

management.

While AI has the potential to transform ASD di-

agnosis and care, several challenges must be addressed.

Key concerns include reducing biases in AI algorithms,

protecting sensitive patient data, and ensuring AI sys-

tems are transparent and easy to understand for health-

care providers.

To overcome these challenges, developing clear

regulations and rigorous testing standards is crucial.

Collaboration among experts in technology, health-

care, ethics, and policy is essential to create eﬀective

and ethically sound solutions. International coopera-

tion and knowledge sharing will also be necessary to

standardize AI applications in ASD care globally.

1.5 Leveraging Technology for Individuals with ASD in India

1.5 Leveraging Technology for

Individuals with ASD in India

In India, innovative technologies are playing a

critical role in addressing the unique challenges faced

by individuals with ASD. For instance, technology is

being leveraged in various ways to support individuals

with ASD.

Personalized E-Learning Platforms

Firstly, personalized e-learning platforms are utiliz-

ing adaptive learning systems to personalize educa-

tional content to individual capabilities. This approach

fosters inclusive learning environments through vir-

tual classrooms and interactive lessons, making educa-

tion more accessible and engaging for individuals with

ASD.

Assistive Technologies for Enhanced Commu-

nication

In addition, assistive technologies powered by AI-driven

speech recognition and audio-visual tools are enhanc-

ing communication for individuals with ASD. These

technologies improve auditory and visual processing,

enabling individuals to better comprehend spoken lan-

guage and follow instructions.

Gaming for Developmental Milestones

Furthermore, customized games are transforming ther-

apy for individuals with ASD by targeting emotional

regulation, cognitive skills, and social interactions.

These engaging tools make therapy enjoyable while

promoting developmental milestones, essential life skills,

and conﬁdence.

Moreover, these technologies promote inclusion,

empowerment, and support cognitive and social devel-

opment, helping individuals with ASD to reach their

full potential. By harnessing the power of technol-

ogy, India is creating a more inclusive and supportive

environment for individuals with ASD.

1.6 Technology-Driven Solutions for

ASD

Furthermore, numerous AI-powered tools are cater-

ing to various needs such as communication, social

skills training, emotional regulation, behavioral analy-

sis, and virtual assistance. These tools address critical

areas essential for individuals with ASD to lead fulﬁll-

ing lives. For example:

Communication and Social Skills

Applications like Avaz and Wysa facilitate ef-

fective communication, while platforms such as

Invention Robot and AHA! provide personalized

therapy plans and social skills training. Conse-

quently, these tools empower individuals with

ASD to navigate social interactions with greater

ease and conﬁdence.

Emotional Support and Regulation

Platforms such as SnehAI and Wysa oﬀer emo-

tional support through advanced natural language

processing (NLP) and machine learning algo-

rithms. Moreover, these tools provide guidance,

helping individuals with ASD manage their emo-

tions and build resilience.

Behavioral Analysis and Early Detection

Tools like Autism Speak India and CARE4AUTISM

utilize AI for behavioral analysis and early de-

tection of ASD-related challenges. As a result,

these platforms oﬀer personalized recommenda-

tions that guide parents and caregivers in provid-

ing timely and eﬀective interventions.

Virtual Assistants

Virtual assistants like Alexa and Google Assis-

tant enhance accessibility and daily task man-

agement for individuals with ASD, promoting

independence and simplifying their routines.

These AI innovations are reshaping ASD manage-

ment in India, oﬀering personalized support and im-

proving outcomes for individuals with ASD and their

families. Moreover, as technological advancements re-

shape ASD care, their impact extends beyond therapy

and diagnosis.

1.7 Empowering Employment: Adult Autism Workstations

By fostering enhanced communication, emotional

regulation, and behavioral management, these inno-

vations prepare individuals with ASD for meaningful

participation in the workforce. Ultimately, the tools

not only empower individuals to develop essential life

skills but also create pathways for inclusivity and soci-

etal integration.

1.7 Empowering Employment: Adult

Autism Workstations

The integration of adults with ASD into the work-

force is a crucial step towards inclusivity and inde-

pendence. Specialized adult autism workstations in

India are designed to leverage the strengths of indi-

viduals with ASD, accommodate their unique needs,

and foster societal acceptance. Several initiatives are

promoting employment opportunities for adults with

ASD, including:

Examples of Adult Autism Workstations in In-

dia

Several initiatives are promoting employment op-

portunities for adults with ASD:

NGO-Led Initiatives: Organizations like Ac-

tion for Autism (AFA), Aarambh India, and Tamana

oﬀer vocational training programs, focusing on

skills like data entry, art, and computer-based

work.

Corporate-Led Programs: Companies like SAP

Labs and Wipro have initiated programs to hire

neurodiverse individuals, oﬀering roles in soft-

ware testing and data management.

Specialized Centers in South India: V-Excel

Educational Trust in Chennai and Sankalp in Hy-

derabad provide training in various skills, pro-

moting ﬁnancial independence.

Breaking Barriers and Promoting Inclusivity

These initiatives not only provide employment op-

portunities but also foster a society that values and ac-

cepts individuals with ASD. By highlighting the talents

of neurodiverse individuals, these programs are:

Dismantling societal stigmas

Promoting a culture of inclusion

Empowering adults with ASD to lead autonomous

lives

Enabling them to contribute meaningfully to their

communities

By integrating technology-driven solutions and

fostering inclusive employment opportunities, India

is making signiﬁcant strides towards building a so-

ciety that supports, empowers, and embraces individ-

uals with ASD. These innovations address gaps in re-

sources, accessibility, and awareness, tailoring solu-

tions to the speciﬁc needs of the population.

References

1. Qin, L., Wang, H., Ning, W. et al. (2024).

New advances in the diagnosis and treatment of

autism spectrum disorders. European Journal

of Medical Research, 29, 322. https://doi.org/

10.1186/s40001-024-01916-2.

2. R

ego, A. C. M. do, & Ara´ujo-Filho, I. (2024).

Artiﬁcial Intelligence in Autism Spectrum Dis-

order: Technological Innovations to Enhance

Quality of Life: A Holistic Review of Current

and Future Applications. International Jour-

nal of Innovative Research in Medical Science,

9(09), 539–552. https://doi.org/10.23958/ijirms/

vol09-i09/1969.

3. Meneses do R

ego, A. C., & Ara´ujo-Filho, I.

(2024). Leveraging Artiﬁcial Intelligence to

enhance the Quality of Life for patients with

Autism Spectrum Disorder: A Comprehensive

Review. European Journal of Clinical Medicine,

5(5), 28–38. https://doi.org/10.24018/clinicmed.

2024.5.5.350.

4. AI-based tools for Autism Spectrum Disorder in

India.

5. Adult Autism Workstations in India.

1.7 Empowering Employment: Adult Autism Workstations

About the Author

Dr. Kalyani Bagri is a Senior Research

Associate at Fernandez Foundation in Hyderabad, In-

dia, where she plays a pivotal role as the lead data sci-

entist in the Neonatology Department. She earned her

Ph.D. in Astrophysics from Pt. Ravishankar Shukla

University, in collaboration with IUCAA and TIFR

Mumbai. In her current role, Dr. Bagri independently

integrates Artiﬁcial Intelligence (AI) into neonatal

care. She is leading critical projects, including the

development of a Bronchopulmonary Dysplasia esti-

mator to guide strategic decisions and optimize steroid

usage in neonates, and a sepsis calculator designed to

detect sepsis in neonates prior to clinical recognition.

Additionally, she is involved in a range of initiatives

aimed at advancing neonatal health outcomes through

her innovative and data-driven approaches.

Part IV

General

BSG-IDE: Revolutionizing Academic

Presentations with AI-Powered Automation

and Real-Time Media Integration

by Ninan Sajeeth Philip

airis4D, Vol.2, No.12, 2024

www.airis4d.com

1.1 Introduction

Traditional LaTeX-based tools have long domi-

nated the landscape of academic presentations. Still, a

groundbreaking open-source new platform, BSG-IDE,

is changing the game with its innovative approach to

slide creation. This sophisticated yet user-friendly in-

tegrated development environment combines artiﬁcial

intelligence, real-time media processing, and intelli-

gent content management in ways previously unseen in

academic presentation software. BSG-IDE is designed

to work across major OS platforms but is tested only in

Linux, the most popular OS for scientiﬁc applications.

1.2 Simplicity is the Key

Simplicity is the key feature that makes BSG-IDE

unique. It combines Beamer, Latex, and Python plat-

forms to create an ecosystem that easily and creatively

translates ideas into visual media. As the GitHub page

of the IDE claims, it is easy to prepare a talk on a

subject where the speaker has expertise, but it is more

challenging to prepare a presentation slide on the same

topic. BSG-IDE is just for bridging this gap.

This article aims to give the reader a walkthrough

experience where, due to space constraints, we may

discuss only some of the salient features the application

provides. BSG-IDE is a product of airis4D, and the

author, Ninan Sajeeth Philip, is the dean of research

at airis4D. The package’s source code is over 10,000

lines in Python and is released on GitHub under the

Creative Commons License. It is Free.

1.3 Installation

BSG-IDE can be installed using pip install in a

Python environment.

1. Make sure that you have installed Python and

Latex on your system. This is a simple procedure

that anyone can do by googling for ﬁve minutes.

2. Create a virtual environment of your choice us-

ing Python by issuing the command “python -m

venv my python” (where my python can be any-

thing that will be your initial Python installation

folder).

3. After creating the virtual environment, we must

activate it with the command:

“source my python/bin/activate”

1.4 Let’s Get Started

and press enter.

4. Make sure that your computer is connected to

the internet and issue the command “pip install

bsg-ide” and that will install the bsg-ide ﬁles

automatically on your system.

5. Finally run “bsg-ide –ﬁx” from a terminal to

ensure that all the packages are installed and

is set up in the OS menu (usually under Oﬃce

Menu).

BSG-IDE is in constant development and at the time

of writing this article, it is version 3.2. To upgrade

your installation to the latest version, you can issue the

command “pip install bsg-ide –upgrade.”

1.4 Let’s Get Started

BSG-IDE control panel is at the bottom. The File

options include the facility to create a New presentation

document, Open an existing document and Save it.

The document is saved as a simple text ﬁle with latex-

embedded commands. This makes it very convenient

to check spelling and grammar on any platform.

1.5 New Presentation

Clicking on the New button generates a clean and

empty screen. The ﬁrst option is to give a proper title

for the talk. This can be done by pressing the Pre-

sentation Settings icon, which opens a clean form for

entering the details. In addition to the standard fea-

tures, there is an option to add a logo to every slide in

the presentation. Fill in the content and Save the set-

tings, completing the initial procedure. All necessary

preamble for the latex ﬁle will be automatically added.

If you want to add or remove any feature from the

preamble (for a latex expert), there is an Edit Preamble

button to view and edit it. At this point, your presen-

tation title page is ready, First, we have to save our ﬁle

by clicking the save button. By default, it might open

in your Documents folder. Give the ﬁle a name. Let us

give it the name Demo.txt, which will save your title

page to the folder. To view it, we need at least one

slide in our presentation. We can create it by clicking

the New Slide button on the left sidebar menu. We

are now ready to preview our presentation. Click the

Generate PDF and then the Preview PDF buttons. You

can see the PDF output, provided you have an existing

PDF reader in your system.

Strangely, you may observe a blank white screen

the same size as the presentation to the right of the slide.

That is the space for your seeker notes. Yes, that’s right.

You have ample space to write all your notes. Not just

text notes, you can use latex commands to organise

them add images, descriptions, and everything with

plain latex commands! For example, suppose I want

to refer to an image to explain the slide, but I do not

want to show the audience. I can add it with a simple

command like:

\includegraphics[width=0.5\textwidth]{Image}

in the Presentation notes window. That image will

appear on my notes page to help me navigate the slides

easily.

What if I don’t want this option? The presentation

1.6 Pympress

Notes panel also have three buttons: Slide Only, Notes

Only and Slides + Notes. Just click one of them and

then click the Generate PDF button and what you get

will be according to your choice. The Notes Only

feature is beneﬁcial if you want to keep a printout of

the notes for reference. It will be just the notes in the

same order as the slides but without taking up space

for the slides.

1.6 Pympress

The PDF preview displayed the slide and the Notes

together on the same screen, which is not one desire.

Thankfully, there is a compelling open-source program

to help us with that. Pympress can split the screen to

display the slide on the projector and the notes on your

laptop, both in full-screen mode. BSG-IDE installs

Pympress during installation, so you do not need to

install it separately. Just click the Presnt with Notes

icon, and that’s it.

1.7 Overleaf Integration

Overleaf is a very powerful latex online platform

with free and premium versions. The greatest ad-

vantage is that it has almost everything you require,

from writing the latex code to generating professional-

quality pdf documents. It also allows one to upload a

latex source ﬁle to generate the pdf ﬁle on the web plat-

form, giving the advantage of sharing and collaborating

in idea development. BSG-IDE has an Export to Over-

leaf icon that can convert your BSG-IDE-developed

ﬁles for use with Overleaf. Because of the huge size

of Latex compilers and support systems, most mobile

systems do not support latex compilation. BSG-IDE is

planning a mobile version, and this integration feature

with Overleaf will be a crucial part.

1.8 The Magic World of BSG-IDE

Camera and Screen Image Capture

Integrating the camera into the IDE is an excel-

lent feature while preparing scientiﬁc presentations, as

most scientiﬁc instruments support a camera interface.

For example, capturing images taken with the micro-

scope or telescope at full resolution directly to the IDE

saves a lot of time. The researcher can share it with his

collaborators after adding his notes as a pdf document.

Some systems do not allow camera integration

but display the images directly on the computer screen.

This is where the screen capture option becomes handy.

As seen in the bottom menu, the screen capture has two

options. One is a single shot of a selected region on

the screen. The screen capture mode is initiated by

clicking the Screen icon in the top menu. The capture

region is deﬁned by holding the mouse’s left button

and drawing a rectangle over the region to be captured.

Releasing the mouse button will capture the screen at

that instant. After each click, the system is ready for the

next selection and capture. Unlike many other screen

capture packages, BSG-IDE does not freeze the screen

during capture; one can thus capture the exact moment

they want to capture by releasing the button.

While single-frame capture is the default option, a

continuous capture mode creates an animation of a se-

quence of frames captured. Selecting the option opens

up a menu where the user can specify the number of

frames and the delay between frames. This time-lapse

image capture feature is handy for scientiﬁc monitor-

ing of the progress of slow processes in biosciences

and chemical processes.

Multimedia Integration

BSG-IDE employ Smart Media Management and

AI-Powered Content Suggestions. Unlike conventional

presentation tools, BSG-IDE incorporates an intelli-

gent media management system beyond simple ﬁle in-

sertion. When users need visual content for their slides,

the system automatically analyses the title and content

to construct relevant search queries. This AI-powered

feature suggests appropriate images and handles the

entire media pipeline - from downloading and opti-

mising content to generating previews and managing

diﬀerent media types. More importantly, it produces an

1.8 The Magic World of BSG-IDE

acknowledgement citation of the source as a footnote

with a hyperlink to the source.

The system’s ability to handle various media sources

is awe-inspiring. It can process YouTube videos, au-

tomatically generate preview frames and manage play-

back integration. For animated content, it intelligently

extracts key frames and creates optimised versions suit-

able for presentations. This level of media intelligence

is unprecedented in academic presentation tools.

Real-Time Syntax Enhancement and

Visual Feedback

BSG-IDE’s most distinctive feature is its real-time

syntax highlighting system. Unlike traditional LaTeX

editors that treat all text equally, BSG-IDE implements

a sophisticated colour-coding scheme that diﬀerenti-

ates between commands, media directives, mathemat-

ical expressions, and special eﬀects. This visual feed-

back system makes it signiﬁcantly easier for users to

spot potential issues and understand their presenta-

tion’s structure at a glance.

The platform introduces novel text eﬀects and vi-

sual enhancements beyond standard LaTeX capabil-

ities. Users can add glowing text, create shadow ef-

fects, and implement gradient-based highlights through

an intuitive interface that automatically generates the

complex underlying LaTeX code. This brings visual

sophistication to academic presentations that were pre-

viously diﬃcult to achieve.

Dual-Screen Presentation Mode with

Integrated Notes

One of BSG-IDE’s most innovative features is its

comprehensive handling of presentation notes. Unlike

traditional systems where notes are an afterthought,

BSG-IDE implements a sophisticated dual-screen pre-

sentation mode that seamlessly integrates speaker notes

with the presentation. The system allows for the rich

formatting of notes, including colour-coding, empha-

sis, and even timing markers - features typically found

in commercial presentation software but rarely avail-

able in academic presentation tools.

The notes system includes innovative templates

for diﬀerent types of content (key points, technical de-

tails, questions and answers). It automatically adjusts

the presentation layout based on the presence and com-

plexity of notes. This level of note integration is unique

among LaTeX-based presentation tools.

Dynamic Layout Engine and Intelligent Content

Positioning The platform’s layout engine represents a

signiﬁcant advancement over traditional beamer pre-

sentations. Instead of rigid templates, BSG-IDE im-

plements a dynamic layout system that automatically

adjusts based on content type and density. It can auto-

matically arrange multiple images in a mosaic pattern,

create picture-in-picture eﬀects, and implement sophis-

ticated overlays without requiring manual positioning

from the user.

This intelligent layout system extends to handling

diﬀerent aspect ratios and screen sizes, automatically

optimising content placement for maximum visibility

and impact. The system can even adjust text size and

positioning based on the complexity of mathematical

expressions or the presence of special eﬀects.

Combining these features - AI-powered content

suggestions, sophisticated media handling, real-time

syntax enhancement, integrated notes management,

and intelligent layout optimisation - makes BSG-IDE a

signiﬁcant leap forward in academic presentation soft-

ware. It bridges the gap between the mathematical

precision of LaTeX and the visual appeal of modern

presentation tools while adding innovative features that

address long-standing pain points in academic presen-

tations.

For researchers and educators who need to create

engaging presentations without sacriﬁcing mathemati-

cal rigour or visual appeal, BSG-IDE represents a new

generation of tools that understand and anticipate their

needs. As academic presentations continue to evolve,

platforms like BSG-IDE are leading the way in mak-

ing sophisticated presentation creation more accessible

and eﬃcient while maintaining the high standards ex-

pected in academic contexts.

1.8 The Magic World of BSG-IDE

About the Author

Professor Ninan Sajeeth Philip is a Vis-

iting Professor at the Inter-University Centre for As-

tronomy and Astrophysics (IUCAA), Pune. He is also

an Adjunct Professor of AI in Applied Medical Sci-

ences [BCMCH, Thiruvalla] and a Senior Advisor for

the Pune Knowledge Cluster (PKC). He is the Dean

and Director of airis4D and has a teaching experience

of 33+ years in Physics. His area of specialisation is

AI and ML.

About airis4D

Artiﬁcial Intelligence Research and Intelligent Systems (airis4D) is an AI and Bio-sciences Research Centre.

The Centre aims to create new knowledge in the ﬁeld of Space Science, Astronomy, Robotics, Agri Science,

Industry, and Biodiversity to bring Progress and Plenitude to the People and the Planet.

Vision

Humanity is in the 4th Industrial Revolution era, which operates on a cyber-physical production system. Cutting-

edge research and development in science and technology to create new knowledge and skills become the key to

the new world economy. Most of the resources for this goal can be harnessed by integrating biological systems

with intelligent computing systems oﬀered by AI. The future survival of humans, animals, and the ecosystem

depends on how eﬃciently the realities and resources are responsibly used for abundance and wellness. Artiﬁcial

intelligence Research and Intelligent Systems pursue this vision and look for the best actions that ensure an

abundant environment and ecosystem for the planet and the people.

Mission Statement

The 4D in airis4D represents the mission to Dream, Design, Develop, and Deploy Knowledge with the ﬁre of

commitment and dedication towards humanity and the ecosystem.

Dream

To promote the unlimited human potential to dream the impossible.

Design

To nurture the human capacity to articulate a dream and logically realise it.

Develop

To assist the talents to materialise a design into a product, a service, a knowledge that beneﬁts the community

and the planet.

Deploy

To realise and educate humanity that a knowledge that is not deployed makes no diﬀerence by its absence.

Campus

Situated in a lush green village campus in Thelliyoor, Kerala, India, airis4D was established under the auspicious

of SEED Foundation (Susthiratha, Environment, Education Development Foundation) a not-for-proﬁt company

for promoting Education, Research. Engineering, Biology, Development, etc.

The whole campus is powered by Solar power and has a rain harvesting facility to provide suﬃcient water supply

for up to three months of drought. The computing facility in the campus is accessible from anywhere through a

dedicated optical ﬁbre internet connectivity 24×7.

There is a freshwater stream that originates from the nearby hills and ﬂows through the middle of the campus.

The campus is a noted habitat for the biodiversity of tropical Fauna and Flora. airis4D carry out periodic and

systematic water quality and species diversity surveys in the region to ensure its richness. It is our pride that

the site has consistently been environment-friendly and rich in biodiversity. airis4D is also growing fruit plants

that can feed birds and provide water bodies to survive the drought.