Closing the reaction-to-process loop with knowledge-intensified AI

Abstract Fine-chemical synthesis is extensively documented, yet kinetic and engineering knowledge for safe operation and scale-up remains scarce. Automated experiments supply additional data without necessarily yielding reusable process knowledge. We develop a knowledge-intensified closed loop (KI-cLoop) integrating chemical knowledge, automated continuous-flow experiments and physics-informed modelling. Its foundation, the Chemical World Skeleton Knowledge Graph (CWSKG), comprises approximately 3.6 million entities and 22 million relations across materials, reactions and processes. CWSKG organizes traceable evidence, experiments generate time-resolved data, an experimentally validated hybrid PINN–ODE model infers kinetics, and reactor models constrain scale-up. Across four strongly exothermic, data-scarce organic peroxide synthesis routes, kinetic fits (R² > 0.92) guided optimization considering yield, selectivity and conversion. Screening predicted a 17.8% relative yield improvement over the best measured yield for one route. Reactor simulations identified a heat-transfer-constrained scale-up boundary with net single-tube production approximately 126 times that of the reference tube. These outputs enrich CWSKG, advancing reaction records into experimentally supported, physically interpretable, scale-dependent process knowledge.