Large Language Models for Physics Instrument Design

TL;DR

The paper assesses whether prompting-based large language models (LLMs) can contribute to physics instrument design by proposing detector layouts under fixed simulation models and comparing them to reinforcement learning (RL) baselines. Without task-specific training, LLMs generate valid, resource-aware configurations that leverage broad detector-design knowledge and particle–matter interactions, achieving substantial gains over baseline designs and recovering a large fraction of RL performance. A minimal hybrid variant—LLM proposals followed by a local trust-region (TR) refinement—further narrows the gap to RL, demonstrating the potential of LLMs as meta-planners that organize design studies and coordinate optimization workflows. The results motivate hybrid pipelines where LLMs structure design hypotheses and RL executes reward-driven optimization, enabling more automated, closed-loop instrument design for future experiments.

Abstract

We study the use of large language models (LLMs) for physics instrument design and compare their performance to reinforcement learning (RL). Using only prompting, LLMs are given task constraints and summaries of prior high-scoring designs and propose complete detector configurations, which we evaluate with the same simulators and reward functions used in RL-based optimization. Although RL yields stronger final designs, we find that modern LLMs consistently generate valid, resource-aware, and physically meaningful configurations that draw on broad pretrained knowledge of detector design principles and particle--matter interactions, despite having no task-specific training. Based on this result, as a first step toward hybrid design workflows, we explore pairing the LLMs with a dedicated trust region optimizer, serving as a precursor to future pipelines in which LLMs propose and structure design hypotheses while RL performs reward-driven optimization. Based on these experiments, we argue that LLMs are well suited as meta-planners: they can design and orchestrate RL-based optimization studies, define search strategies, and coordinate multiple interacting components within a unified workflow. In doing so, they point toward automated, closed-loop instrument design in which much of the human effort required to structure and supervise optimization can be reduced.

Figures (11)

• Figure 1: Calorimeter performance across 350 design iterations. Energy resolution versus design iteration for multiple LLMs (GPT-OSS-20B, GPT-OSS-120B, GPT-5, and Gemini 2.5 Pro) over 350 iterations. The panels correspond to electromagnetic showers at 50 GeV (top left) and 100 GeV (top right), and hadronic showers at 50 GeV (bottom left) and 100 GeV (bottom right).
• Figure 2: Calorimeter scalar reward as a function of design iteration for multiple reconstruction models (GPT-OSS-20B, GPT-OSS-120B, GPT-5, and Gemini 2.5 Pro), shown over 350 optimization iterations.
• Figure 3: Calorimeter best design found during different intervals during the training process for a standalone LLM (GPT-OSS-20B) on the left and a LLM (Gemini 2.5 Pro) + TR refinement on the right.
• Figure 4: Spectrometer performance across 350 design iterations. Tracking efficiency (left panels) and momentum resolution (right panels) as a function of design iteration over 350 proposal steps. The top row corresponds to 10 GeV tracks, and the bottom row to 100 GeV tracks. Curves are shown for the evaluated language models (GPT-OSS-20B, GPT-OSS-120B, GPT-5, and Gemini 2.5 Pro).
• Figure 5: Spectrometer scalar reward as a function of design iteration over 350 proposal steps. The reward aggregates the spectrometer performance metrics into a single objective. Curves are shown for the evaluated language models (GPT-OSS-20B, GPT-OSS-120B, GPT-5, and Gemini 2.5 Pro).