Formulation screening stops guessing: KolTa folds a temperature physics residual into acquisition, cutting drug-solubility queries by about 18%
Synopsis
For formulation discovery with few measurements and a fixed experimental budget, KolTa splits property prediction into component base contributions, role-conditioned interactions and an anchored temperature response, trained with a derivative-based physics constraint. Its acquisition rule calibrates ensemble disagreement together with the physics residual at the target temperature against out-of-fold errors. Across binary-mixture viscosity, drug solubility, ionic-liquid viscosity and simulated electrolyte conductivity, mean AULC ranks first or second on every task, and on the drug task the median queries to reach 80% recovery are about 18% fewer than the fastest comparator.
Interpretation
Formulation discovery can now prioritize the top 5% of candidates at a target temperature under a fixed query budget, rather than only predicting properties more accurately across a batch. Earlier mixture active-learning workflows mostly used expected improvement or greedy rules, and temperature was usually concatenated into the mixture representation without an explicit thermodynamic constraint on temperature derivatives. On three measured tasks and one simulated task, KolTa ranks first or second in mean AULC on every task, and in comparisons paired by task and seed it wins 75 of 78 pairs against seven of the eight comparator strategies.
The predictor encodes each component's molecular graph, sums base contributions by component role, runs one round of directed message passing for interactions, and describes temperature response with a network anchored at a reference temperature. The temperature subnetwork's output at the reference temperature is subtracted, so it predicts only the change relative to that temperature, while the physics residual compares the network derivative with the derivative given by Arrhenius, van't Hoff or VFT relations. The physics loss is computed at five fixed temperatures per task without requiring measurements there; acquisition weights are refitted each round from five-fold out-of-fold signed errors, and the physics residual receives positive weight in 115 of 120 rounds.
Adding the physics residual to the acquisition margin raises mean AULC over the disagreement-only calibrated variant on both ablation tasks. The disagreement-only variant A1 reaches mean AULC of 0.729 on the drug task and 0.736 on the ionic-liquid task. Full PRC-UCB raises these to 0.738 and 0.748 and wins five of six paired runs; on the drug task the median queries to 80% macro recovery are 348, versus 424 for the fastest comparator, SolvGNN-UCB.
Perspective
The evaluation is retrospective: on four tasks, NIST binary-mixture viscosity, solubility across 86 drug groups, ILThermo ionic-liquid viscosity and DiffMix simulated electrolyte conductivity, strategies are compared by discovery curves and locked-set error under a fixed query budget. The physics loss is applied at five fixed temperatures per task, and temperature is the model's only varying condition.
Extrapolation beyond the training temperature range has not been evaluated, and writing physical relations directly into the temperature-response function is only a direction the authors propose. The model currently treats temperature as its only varying condition; extending it to jointly varying temperature and pressure would require additional condition inputs and coupled physical constraints. Code will be made publicly available upon acceptance.