Direct discovery · Materials science · Chemistry
SynAgent revises thin-film growth hypotheses through an 18-experiment physical campaign
LiCoO2 films and an experiment-linked growth map identifying a 650–690 °C high-crystallinity window.
Summary
SynAgent synthesized and characterized 18 LiCoO2 films while alternating tests expected to support or contradict its hypotheses. Failed predictions redirected the campaign toward a 650–690 °C growth window. It generated two reusable analysis routines and an explicit account of the temperature-dependent observations. The optimum was also found in earlier Bayesian-optimization work on the same platform.
AI role
GPT-5.5 agents selected deposition temperatures, generated XRD and microscopy analysis code, and revised hypotheses after physical measurements.
Narrative role
Adds physical evidence of an agent revising a process hypothesis after contradictory experiments, rather than merely proposing a candidate or optimizing a simulated objective.
Caveat
One material and one controllable variable, with each condition tested once. Analysis code was self-tested and support/refutation judged by the LLM. No matched-budget comparison with Bayesian optimization or independent validation of mechanistic explanations.