From Small-Data Conductivity Prediction to Bayesian Optimization of Electrolyte Formulations with a Tabular Foundation Model

We evaluate whether a frozen tabular foundation model, TabPFN, can use limited measurements to predict conductivity and guide electrolyte formulation search without explicit transport equations or task-specific physics-constrained training. On the AMPERE conductivity dataset, 40 labeled context samples yield a mean absolute percentage error (MAPE) of 5.56% on 60 held-out measurements, close to the 5.32% reported for the original physics-augmented AMPERE model and below the conventional regression baselines under locally matched conditions. Additional evaluations on CALiSol-23, Flores, ionic-liquid, and polymer-electrolyte datasets support broader predictive applicability, while target-domain measurements improve shifted-domain predictions through context updates alone. The main Bayesian optimization application revisits the AMPERE rate-performance search setting: TabPFN-predicted conductivity and operating salt concentration form a two-dimensional representation for a separate Gaussian process (GP) surrogate. Across two temperatures, measured-pool replays require 13.90 evaluations on average to find the best measured formulation, reducing cost by 45.6% relative to random search and 9.1% relative to composition inputs. Two further applications examine TabPFN as a direct objective surrogate. In sequential Clio conductivity replay, a fixed acquisition rule requires 11.5 evaluations on ten new starts, 38.2% fewer than random search and close to independently selected GP (11.0). In the ionic-polymer study, seven initial measurements support ranking the source study's recommended formulation at 21–27 among 497 candidates, versus 44 for Gaussian process regression. Together, these cases show how context-based prediction can support candidate selection through property-informed representations and direct objective surrogates.