Mixed campaignchemistry · solid-state chemistry and battery materials
A-Lab GPSS runs 352 spinel-electrolyte experiments
A GPT-5 agent proposed and executed 263 samples in a 352-experiment lithium-halide campaign, raising the joint phase-purity and conductivity hit rate while preserving the full low-yield denominator.
Editorial context
What happened
Humans designed the first 77 samples and 12 later samples. The agent proposed and executed 263 samples while using abnormality detection, inductive pattern finding, Bayesian optimization, and a reflection memory. The public experiment table is more informative than a success label: it covers 19 metals and 124 of 171 possible pairwise metal combinations, and it retains every unsuccessful sample alongside the improving hit rate.
Keep in view
Humans constrained the search space and handled several physical and analytical steps. The final joint hit rate was 5.33%, so the result is an optimization campaign with many unsuccessful experiments rather than one autonomous breakthrough.
Chronology
Timeline
The 53-day laboratory campaign beganattempt · day date
Campaign ended after 352 synthesized samplesattempt · day date
Paper, prompts, code, and campaign data releasedpreprint · day date
Trust boundaries
Validation
All proposed samples were physically synthesized and evaluated using X-ray diffraction and electrochemical impedance spectroscopy.
The team later updated 13 conductivity values and human reinspection found occasional phase-fitting problems.
System
AI and tools
- A-Lab GPSS
- GPT-5 with high reasoning effort
- Bayesian optimization
People
Human contributors
- Yuxing Fei
- Bernardus Rendy
- Gerbrand Ceder
- A-Lab GPSS collaborators
Organizations
Affiliations
None recorded.
Primary materials