Machine learning unlocks robust convergence for chemical process simulations

Convergence failures in process simulation are a critical barrier to designing the novel chemical processes required for a sustainable economy, often leaving engineers unable to distinguish numerical instability from true physical infeasibility. Here we introduce COSMIC, a machine-learning copilot that reframes convergence as a surrogate-based optimization problem. Numerical examples show that COSMIC can converge industrial flowsheets where conventional solvers may fail. This enhanced reliability can significantly expand the volume of explorable design space by more than 200% relative to conventional solver’s feasible region. By transforming convergence from a fragile bottleneck into a systematic, data-driven step, COSMIC enables reliable design space exploration and accelerates the innovation of next-generation chemical processes.