Feb 26, 2026
Solid electrolytes as an autonomous-lab benchmark
Why lithium solid electrolytes are a good testbed for closed-loop AI: the objectives are clear, the constraints conflict, and synthesis remains decisive.
Why This Benchmark Is Hard Enough
Solid electrolytes make a useful autonomous-lab benchmark because there is no single magic metric. Ionic conductivity matters, but so do electrochemical stability, phase purity, processability, mechanical behavior, abundance, and interface compatibility. A candidate that wins one column can fail badly in another.
That makes the problem closer to real energy research than a clean leaderboard. The discovery system has to decide whether to exploit a known chemistry family, explore a surprising substitution, or run a diagnostic synthesis that tests whether the model is overconfident.
The Loop
The loop begins with a generator that proposes compositions constrained by charge balance, known crystal families, supply-chain flags, and estimated stability. A fast scoring layer filters candidates into synthesis batches. Measurements then update both the candidate model and the experiment planner.
The planner is the important part. It should not simply choose the candidates with the highest predicted conductivity. It should choose the batch that best improves the lab state: a mix of promising candidates, uncertainty reducers, and controls that reveal whether the synthesis route itself is failing.
What Counts As Progress
Progress is not only a new electrolyte. It can be a better rule for excluding unstable substitutions, a synthesis protocol that improves phase purity, or a measurement that exposes a bad descriptor. Automated discovery should make these intermediate gains visible.
We score each campaign by learning efficiency: how much the posterior candidate map improves per synthesis attempt. That discourages shallow optimization and rewards experiments that make the next round better.
Connecting Back To Deployment
A lab discovery does not become an energy technology until it survives scaling, interfaces, packaging, and cost. The benchmark therefore includes deployment tags from the start. Candidates can be marked as spectroscopy targets, synthesis targets, interface targets, or long-horizon speculative targets.
This prevents the system from pretending that every stable computed material is equally useful. The output should be a portfolio of next actions, not a press-release list of discoveries.