GRACE: an Agentic AI for Particle Physics Experiment Design and Simulation

TL;DR

This work establishes experimental design as a constrained search problem under physical law and introduces a new benchmark for autonomous, simulation-driven scientific reasoning in complex instruments.

Abstract

We present GRACE, a simulation-native agent for autonomous experimental design in high-energy and nuclear physics. Given multimodal input in the form of a natural-language prompt or a published experimental paper, the agent extracts a structured representation of the experiment, constructs a runnable toy simulation, and autonomously explores design modifications using first-principles Monte Carlo methods. Unlike agentic systems focused on operational control or execution of predefined procedures, GRACE addresses the upstream problem of experimental design: proposing non-obvious modifications to detector geometry, materials, and configurations that improve physics performance under physical and practical constraints. The agent evaluates candidate designs through repeated simulation, physics-motivated utility functions, and budget-aware escalation from fast parametric models to full Geant4 simulations, while maintaining strict reproducibility and provenance tracking. We demonstrate the framework on historical experimental setups, showing that the agent can identify optimization directions that align with known upgrade priorities, using only baseline simulation inputs. We also conducted a benchmark in which the agent identified the setup and proposed improvements from a suite of natural language prompts, with some supplied with a relevant physics research paper, of varying high energy physics (HEP) problem settings. This work establishes experimental design as a constrained search problem under physical law and introduces a new benchmark for autonomous, simulation-driven scientific reasoning in complex instruments.

Figures (12)

• Figure 1: Architecture overview of GRACE. Natural language prompts or papers are parsed by the Task Classifier, which feeds the Workflow Planner. The Tool Router dispatches steps to physics tools. Outputs are validated by the Physics Verifier against the Knowledge Graph. The Self-Evaluator and Error Recovery modules enable adaptive re-planning when verification fails.
• Figure 2: The GRACE control loop. Each iteration begins with observing the current experiment state and querying the Knowledge Graph, proceeds through planning and execution, validates outputs against physics constraints, updates provenance and memory, and iterates with refined parameters or escalated fidelity.
• Figure 3: Photoelectron yield distributions at 0.1 MeV (left), 1.0 MeV (center), and 5.0 MeV (right). The distributions demonstrate linear scaling of light yield with energy and improving relative resolution at higher energies consistent with Poisson statistics.
• Figure 4: Nuclear recoil discrimination analysis. Left: S1 signal distributions for electronic recoils (blue) and nuclear recoils (red), demonstrating clear separation for signal-background discrimination. Right: Quenching factor (nuclear/electronic light yield ratio) as a function of recoil energy, showing convergence toward unity at higher energies.
• Figure 5: Spatial response characterization. Top row: Position resolution versus energy for 3D/radial (left) and X/Y/Z components (right). Bottom row: Spatial uniformity map showing relative response across the detector cross-section (left) and position resolution as a function of radial position (right).