Bilegsaikhan Tuvshintulga

Degree subject: Computer Science with a pathway in Data Science

LinkedIn profile: https://www.linkedin.com/in/bilegsaikhan-tuvshintulga

What interested you in your degree and what do you like about it?

I chose this degree because I’ve always been drawn to problem-solving and to understanding how intelligent systems work. I’m particularly interested in machine learning, research workflows and computational tools that support scientific discovery.

Why did you apply for this internship and what do you hope to gain in completing it?

I applied because I want to build deeper technical experience in machine learning and learn how real research workflows operate. This project combines AI, modelling and HPC in a way that directly fits what I want to develop next. It’s a chance to contribute to meaningful work while learning from experienced researchers.

I hope to gain practical experience with research-grade machine learning workflows, especially working with HPC tools, reproducible experiments and symbolic modelling. I also want to learn how researchers structure complex problems and turn ideas into tested results. The aim is to finish with stronger technical depth and a clearer understanding of how ML is applied in real scientific work.

This project adapts a state-of-the-art symbolic transformer model, originally developed by Alfarano et al. (NeurIPS 2024), to discover Lyapunov functions for real power system models. Lyapunov functions are mathematical certificates that prove a system is stable. Finding them for complex, non-polynomial power systems is an open problem with direct relevance to a reliable and resilient net-zero power grid. The internship delivered three main outputs: a Python package that encodes real power system models in a format compatible with the model's training pipeline, an end-to-end automated training infrastructure running on the University of Liverpool's Barkla HPC cluster, and a replicated baseline model trained to convergence on non-polynomial benchmark datasets. The infrastructure is fully documented and reproducible, designed for handover to the research team to continue fine-tuning on real power system data.

What are the key results of your project?

1. A symbolic transformer model was replicated and trained to convergence on non-polynomial benchmark datasets, achieving 98.5% accuracy on the Lyapunov test set and 76.5% on the Barrier test set. All outputs produced were mathematically valid.
2. Real power system models were encoded in Python using symbolic mathematics and integrated into the training pipeline.
3. An end-to-end automated pipeline was built on the University of Liverpool's Barkla HPC cluster, enabling data generation, model training, evaluation, and environment setup to be run reproducibly by the research team.
4. A fully documented prototype was delivered and handed over, providing the foundation for the next phase: fine-tuning the model on real power system data through parameter perturbation and LMI-based dataset expansion.

How do you feel you have benefitted from completing this internship?

This internship gave me hands-on experience across a range of skills I had not expected to develop at this stage. From working with high-performance computing infrastructure, contributing to a real research codebase, and building reproducible software for a research team. It also gave me a clearer sense of how research software engineering works in practice, which is something you cannot get from coursework alone.

It has directly influenced my thinking about future careers. The N8 CIR programme introduced me to the Research Technology Professional pathway, and this experience has made me seriously consider pursuing a career in that direction, whether as a Research Software Engineer or in a related research infrastructure role.