PC-MCMC-CIGP: New AI workflow discovers chemical reactions with 12.5% yield boost

Extracting interpretable governing equations from sparse, noisy chemical time-series data has been a longstanding challenge because discrete reaction topology and continuous kinetic parameters are tightly coupled. The authors present PC-MCMC-CIGP, a gray-box workflow that synergizes physically constrained MCMC sampling with chemical-informed Gaussian processes. The workflow uses spike-and-slab topology sampling to handle model selection, enforces conservation laws and thermodynamic feasibility, and employs a CIGP residual model for parameter calibration and experimental design. The key contribution is the integration of these components into a unified, reproducible framework with explicit uncertainty-aware acquisition strategies, rather than a novel MCMC or GP method in isolation.

On the H2+Br2 benchmark, the constrained sampler correctly distinguishes elementary radical pathways from deceptive phenomenological fits, demonstrating its ability to uncover true reaction mechanisms. For styrene epoxidation, the CIGP optimization loop achieves a 12.5% improvement in final yield compared to a reported GP-BO baseline. A comprehensive 10-seed acquisition study compares EI, GWU, PC-EI, uncertainty sampling, discrepancy hunting, and random search, revealing different trade-offs: PC-EI substantially reduces low-yield BO suggestions, while EI-style criteria yield the strongest final performance. The work promises to accelerate reaction network discovery and experimental design in chemistry.