AI-Driven Physics Instrument Design: From Reinforcement Learning to Large Language Models

About

Designing modern physics detectors is a high-dimensional, combinatorial problem that has traditionally relied on expert intuition and hand-tuned optimization. This seminar presents two complementary machine-learning approaches that automate this process: reinforcement learning (RL) and large language models (LLMs). Reinforcement Learning agents can explore complex design spaces without fixed parameterizations, producing competitive detector layouts for tasks such as calorimeter segmentation and tracker placement. Pretrained Large Language Models can generate valid detector configurations under the same simulation and reward framework. While RL achieves stronger optimization performance, LLMs reliably produce feasible, resource-aware designs and serve as high-level planners that can guide or structure the search. Together, these works point toward hybrid, closed-loop workflows that combine conceptual planning by LLMs with fine-grained optimization by RL to accelerate the design of next-generation physics instruments.

Bio: Shah Rukh Qasim is a researcher jointly affiliated with the Physik-Institut and the Department of Mathematical Modeling and Machine Learning at the University of Zurich. With a background in computer science, he previously worked on particle reconstruction in high-energy physics during his Ph.D. at CERN with dynamic graph neural networks. His current work focuses on applying machine learning—particularly reinforcement learning—to detector and system design in particle physics, including a core role in the design optimization of the muon shield for the SHiP experiment, alongside broader interdisciplinary ML-driven research projects in epidemiology and supply chains.

coffee will be served at 10h30

00:00:00 Slide 1

00:00:32 Slide 2

00:01:17 Slide 3

00:01:39 Slide 4

00:02:52 Slide 5

00:03:43 Slide 6

00:05:03 Slide 7

00:05:20 Slide 8

00:06:02 Slide 9