Author ORCID Identifier

https://orcid.org/0009-0008-5913-8959

Author Linked-In Account

www.linkedin.com/in/priyanka-priyan-0a2b12147

Biosketch

Dr. M. Priyanka is an accomplished academic and researcher in the field of Computer Science and Engineering, with strong expertise in Machine Learning and Deep Learning. She began her academic journey by completing her Bachelor of Engineering in Computer Science and Engineering at Anjalai Ammal Mahalingam Engineering College, Kovilvenni, Tamil Nadu, graduating with First Class with Distinction. Her performance at the undergraduate level reflects her consistent academic focus and technical capability.

She continued to excel during her Master of Engineering in Computer Science and Engineering at A.R.J. College of Engineering and Technology, Edayarnatham, Tamil Nadu, where she again secured First Class with Distinction. She earned the notable achievement of securing the 7th rank in the Anna University merit list for M.E., underscoring her analytical skills and commitment to academic excellence.

Driven by a deep interest in intelligent systems and data-driven research, she pursued and completed her Ph.D. in Computer Science and Engineering at the Srinivasa Ramanujan Centre, SASTRA Deemed University, Kumbakonam. Her doctoral work focused on advanced techniques in Machine Learning and Deep Learning, contributing to innovations in predictive modeling, data analysis, and AI-driven solutions.

Her research interests include artificial intelligence, neural network architectures, intelligent decision-making systems, and practical applications of machine learning across domains. She is passionate about advancing impactful research, participating in collaborative projects, and supporting learners in emerging technology areas. With strong academic accomplishments and a sustained interest in modern computational methods, M. Priyanka continues to explore and contribute to evolving directions in AI and computing.

Date of Award

17-8-2025

Document Type

Thesis

School

School of Computing

Programme

Ph.D.-Doctoral of Philosophy

First Advisor

Dr.J.Sangeetha

Keywords

Plasma Disruption Prediction; Aditya Tokamak; Active Learning; Long Short-Term Memory (LSTM); Ensemble Learning

Abstract

Tokamaks are nuclear fusion reactors designed to generate sustainable energy by confining plasma, yet plasma disruptions remain a major obstacle as they can damage reactor components and interrupt fusion reactions. Addressing this challenge requires reliable models that can classify plasma discharges and predict disruptions in advance. This thesis develops machine learning and deep learning approaches to improve both classification and forecasting.The first contribution is a semi-supervised active learning framework, Nearest Margin-Ranked Batch Mode Active Learning (NM-RBMAL), integrated with an ensemble model.

Unlike traditional classifiers that operate on static training data, this approach continuously adapts to new plasma conditions, reducing model ageing. Tested on the Aditya dataset of 162 shots, it achieves 94% accuracy with a 93% ROC score. Building on this, a Double-Phase Stacking with Active Learning (DPST-PAL) model combines outputs from multiple classifiers in two phases to enhance robustness and adaptability. This method improves classification performance, achieving 98% accuracy and a 99% F1-score.

For disruption prediction, a Bi-LSTM with Dynamic Time Window Aggregation (Bi-LSTM-DTWA) is proposed to overcome the problem of premature alarms. By dynamically adjusting the time window based on evolving plasma signals, the model provides early and reliable forecasts, predicting disruptions 10–23 ms in advance with low computational cost on 220 Aditya shots. In addition, an unsupervised approach, the Gated Recurrent Neural Network with Dynamic Threshold-based Temporal Differentiation (GRNN-DTTD), is developed to eliminate reliance on fixed thresholds and labeled data. By analyzing temporal fluctuations in plasma current, it achieves 98.9% prediction accuracy with lead times of 12–30 ms.

Together, these models form a comprehensive framework for plasma discharge classification and disruption prediction. They not only enhance accuracy but also adapt to evolving plasma behavior, offering timely warnings with reduced false alarms. The findings contribute toward building reliable predictive systems for tokamak reactors, which is essential for achieving stable and sustainable fusion energy.

Download

Share

COinS