XTQSPRKinetic: A Mixture-of-Experts Framework for Reaction-Space Clustering and Property Learning from SMILES

Abstract

Chemical heterogeneity can limit the ability of a single global structure–property model to represent modified Arrhenius parameters across a reaction network. XTQSPRKinetic combines learned reaction grouping with specialist regression in a configurable mixture-of-experts (MoE) architecture evaluated on a Reaction Mechanism Generator (RMG)-derived n-hexane pyrolysis/steam-cracking network containing 1215 elementary reactions. Using 23 Boruta-confirmed descriptors, a seven-family gate combining Latent Dirichlet Allocation with fuzzy C-means clustering and hard routing to Extreme Gradient Boosting (XGBoost) specialists yielded fitted-network root-mean-square errors (RMSEs) of 7.904, 0.985, and 4.855 for ln A, n, and Ea (kcal mol–1), respectively. Under the same descriptor basis, the learned-family configuration achieved lower reconstruction errors than both the unpartitioned global baseline (corresponding RMSEs of 14.394, 1.819, and 8.946) and size-matched random partitions (mean RMSEs of 10.746, 1.329, and 5.932). RMG-label specialists achieved still lower RMSEs of 5.629, 0.681, and 3.935. This comparison uses 15 RMG classes versus seven learned groups and differs in class composition, partition resolution, and specialist support; the performance gap cannot therefore be attributed to label semantics alone. Learned grouping provides a complementary organization without requiring RMG labels for gate fitting or routing. Residual diagnostics identified rare 1,3-sigmatropic rearrangements among the largest errors, while SHapley Additive exPlanations (SHAP) revealed family dependent use of the shared descriptors. These metrics characterize fitted-network reconstruction; predictive performance on unseen reactions requires separate held-out evaluation.