Researchers at Oak Ridge National Laboratory used a neutron-ray-tracing simulator to train an AI agent to tune virtual instrument settings. On one fixed-layout task, reinforcement learning raised an 8-billion-parameter model’s pass rate from 11.3% to 76.7% on held-out cases. The study does not show that AI designed or operated a physical instrument.
Neutron instruments guide particles from a source through components to study materials. Testing design changes on real equipment is costly, so researchers use simulations to compare candidate settings before experiments. The Oak Ridge team’s NeutronGym turns part of this work into tasks for language-model agents.
An agent can use 22 validated tools to specify an instrument, while McStas simulates neutron paths through it. Four checks score each design: whether it is valid, runs, has the expected structure and meets a physics target. The system measures the result in the simulation rather than asking another language model to judge it.
For the task behind the main result, each case had a hidden reference design that set target beam width and spread in neutron direction. The reinforcement-learning process used NeutronGym’s reward ladder to give partial credit as an agent moved from valid settings and a running simulation toward matching those targets; feedback was not limited to a right-or-wrong score. On 300 held-out cases, Qwen3-8B’s pass rate rose from 11.3% before training to 76.7%. A second training run scored 69%. These were unseen cases within the same task family, not new instrument types.
A separate test shows what remains hard. McStasBench covers 16 tasks derived from 12 published instruments. The best-scoring model rebuilt seven tasks in a one-shot test and five through the tool loop. No model reached the benchmark’s improvement goal: achieving 80% of the gain from a supplied baseline design to a classical-search optimum. A basic reproduction pass confirms beam delivery at named monitors, not an instrument’s defining performance, such as a diffractometer’s resolution.
On a shared 100-case slice of the training task, zero-shot frontier models passed 98–99 cases, compared with 76 for the trained Qwen model. A conventional parameter search, seeded with a mathematical optics solution, reached 81% at the same simulation budget; that result was not statistically distinguishable from Qwen’s 76.7% at the study’s sample size.
NeutronGym is an early test environment for physics-checked AI design, not evidence of autonomous instrument engineering. Its scored tasks keep the layout fixed and adjust only two or three parameters; the benchmark is too small to rank models, and no result has been validated on physical hardware. Broader task families, repeated training runs and tests against real instruments would show whether the approach transfers beyond controlled simulations.
Sources
- NeutronGym paper (full text)
- McStas project
- Contextual image: experimental hall at the ISIS neutron source, by wurzeller, via Wikimedia Commons; used unchanged under CC BY-SA 3.0. It does not depict NeutronGym or Oak Ridge hardware.
