A physics-informed modular neural network for yield prediction and operational optimization in industrial hydrocracking

Efficient operation of industrial hydrocracking units requires simultaneously maximizing middle-distillate yield, minimizing hydrogen consumption, and enforcing thermal safety constraints across multi-bed reactors processing variable feedstocks. Conventional black-box data-driven surrogates lack the internal physical structure needed to produce thermally feasible predictions under changing operating conditions, limiting their utility for refinery decision support. This work presents a Physics-Informed Modular Neural Network (PI-MNN) for surrogate modeling of a commercial single-stage hydrocracking unit, trained and evaluated on over 12,500 h of industrial plant data. The architecture mirrors the physical topology of the process, decomposing the reactor section into sequential bed-level neural modules connected by deterministic quench mixing layers that enforce inter-bed energy conservation by construction. On a chronological holdout evaluation against a monolithic neural network baseline, the PI-MNN achieves MAE reductions of 23.5% for diesel and 15.3% for gasoline yield, while reproducing bed inlet temperatures with accuracy sufficient for industrial reactor monitoring − a capability structurally inaccessible to black-box alternatives. Applied to surrogate-assisted multi-objective optimization, the model resolves bed-level thermal constraint activation at the Pareto boundary and identifies a 24% reduction in make-up hydrogen flow at the balanced optimum, directly translatable to improved unit-level sustainability without capital investment. The results establish that embedding reactor topology directly into the computational graph improves both predictive performance under a genuine shift in the joint operating regime between training and test periods and the physical interpretability of internal model states.