19–23 May 2025
CERN
Europe/Zurich timezone

Learning to Design Physics Instruments for Varying Budgets with Reinforcement Learning

Not scheduled
20m
61/1-201 - Pas perdus - Not a meeting room - (CERN)

61/1-201 - Pas perdus - Not a meeting room -

CERN

10
Show room on map
Poster 6 ML infrastructure: Hardware and software for ML/MLOps Poster Session

Speaker

Sara Zoccheddu (University of Zurich (CH))

Description

The design of advanced physics instruments is a complex and resource‐intensive task—one that requires optimizing numerous parameters (such as the sizes and shapes of various elements) to achieve high performance while meeting stringent cost, material, and spatial constraints. Our new work extends the approach presented in arXiv:2412.10237, which leverages Reinforcement Learning (RL) for instrument design, by further incorporating a varying budget. In this new method, the RL agent does not learn a single detector strategy; instead, it learns several models—each conditioned on a different resource budget. Demonstrated through calorimeter design, this conditioning enables the agent to tailor sensor placement, layer thickness, and overall detector architecture in accordance with available resources in a non-linear fashion. As a result, the algorithm outputs multiple optimized designs corresponding to distinct budgets, thereby allowing decision-makers to assess trade-offs between performance (for example, energy resolution and efficiency) and cost or material consumption. We argue that this approach will prove highly beneficial in the context of high energy physics, as budget constraints during experiment design evolve in response to achievable performance and policy changes.

Would you like to be considered for an oral presentation? Yes

Author

Sara Zoccheddu (University of Zurich (CH))

Co-authors

Nicola Serra (University of Zurich (CH)) Shah Rukh Qasim (University of Zurich (CH))

Presentation materials

There are no materials yet.