Autonomous experiments that update their own physical theory while running
bravo_abad · x · 2026-08-18
Slautin et al. (ACS Nano) propose Bayesian conavigation: a closed-loop framework where an automated experiment and a computational physics model are explored simultaneously.
- Three coupled loops: one explores the experiment, one the physical simulation, and an outer Bayesian loop continuously adjusts the physical model's parameters to reduce theory-experiment mismatch — going beyond standard Bayesian optimization.
- Setup: automated AFM measures surface roughness across a (CrTaWV)x−Mo(1−x) thin-film library while a kinetic Monte Carlo model simulates the same observable; Gaussian-process surrogates guide both spaces and the outer loop updates effective bond energies in the kMC model.
- Results: the model becomes progressively material-specific; the workflow reconstructs the observed roughness maximum near x≈0.7 in 2 hours vs 16+ hours for exhaustive AFM mapping.
- Mechanism: inferred EAB > EAA, EBB — stronger mixed CrTaWV–Mo bonding suppresses surface diffusion at intermediate compositions. The authors caution the fitted parameters may not be unique.
Key AI-for-Science lesson: autonomous discovery needn't choose between black-box optimization and fixed theory — experiment and theory can co-evolve, moving self-driving labs toward refining both experiments and scientific models as they learn.
More from Research
- Sentence Transformers Adds Support for Late Interaction Models — ariG23498 · 2026-08-18
- Paper Clarifies Concepts of LLM Chain-of-Thought Faithfulness — benno_krojer · 2026-08-18
- Prof. teaches AI architectures via hand calculation: Transformer to Mamba — techNmak · 2026-08-18
- Video Model Evaluation Guide: Quantifying Generation Quality — Majumdar_Ani · 2026-08-18
- Study: LLMs as synthetic survey respondents are plausible but not valid — Mantas Lukauskas · 2026-08-18
- Adaption AI Launches Custom Evals for Pro Users — sarahookr · 2026-08-18