SPARC lets experimental evidence rewrite model weights: 8 of 11 recommended conditions fall below 300 mV overpotential
Synopsis
Closed-loop experiments usually fix the model family and representation first, then let new data update only predictions. SPARC builds an evidence matrix from the out-of-sample predictions of competing regression models, assigns each model's influence through a sparse posterior, and combines candidate-dependent predictive dispersion with a residual scale into a trust score that modulates exploration. In an automated alkaline oxygen-evolution campaign, 8 of 11 recommendations fixed in advance reached overpotentials below 300 mV, versus 18 of 125 initial experiments, with a best value of 242 mV.
Interpretation
In materials-discovery loops, the relative influence of competing model families is now revised as experimental evidence accumulates, not only the predicted values. Earlier closed-loop methods typically fixed the representation, model family and exploration assumptions before optimization began, so new measurements updated predictions while other plausible models had little chance to be tested. Across eight published molecular, polymeric and inorganic materials datasets, SPARC reached higher R² than the corresponding literature-reported values on six, with relative gains up to about 77%, without the density-functional-theory- or experiment-derived descriptors used by several published pipelines.
Each candidate model first predicts observations it never saw, and non-negative weights are then allocated by those out-of-sample contributions. Every row of the evidence matrix is a held-out observation and every column a candidate model, so each entry comes from a fit that excluded that response; stability screening keeps a reproducible active set, and a hierarchical half-Cauchy posterior composition sets the weights. Predictive benchmarks were evaluated under nested outer leave-one-out validation, repeating preprocessing, model selection and posterior fitting for each held-out observation; the automated oxygen-evolution campaign used a heterogeneous bootstrap ensemble with a trust-modulated lower confidence bound to recommend 11 conditions from 125 initial experiments.
Perspective
The demonstrated scope covers predictive benchmarks on eight published materials datasets, two finite-pool retrospective searches, two controlled extrapolation tasks, and one prospective automated alkaline oxygen-evolution campaign starting from 125 experiments; recommendations and verification groups were fixed before their outcomes were measured.
Each finite-pool discovery count comes from a single trajectory started from one initial set, so these are case studies; trust is a task- and scale-specific decision score, not a calibrated probability of prediction correctness or measurement error; controlled extrapolation tests expansion along the same experimental variables rather than transfer across chemistries; measurements prioritized for repetition should be read as requiring more evidence, not as shown to be wrong.
