Hybrid modeling of an industrial LPG debutanizer using a differentiable first-principles distillation solver with real plant data

Accurate simulation of industrial distillation columns depends on precise thermodynamic parameters, but these are difficult to obtain when feed streams include uncharacterized species. In natural gas processing, plant gas chromatographs identify only a limited set of key species. Unresolved species are lumped into a single fraction and assigned to the properties of a representative molecule. Since the true composition of this fraction varies with crude source and upstream processing, but the assigned properties remain fixed, systematic uncertainty is introduced into the column model. First-principles models cannot resolve this issue because fixed properties do not reflect actual feed variations, while purely data-driven models lack interpretability and extrapolation capability. This work introduces a hybrid approach for an industrial LPG debutanizer at the Ghana National Gas Company. It combines a Peng–Robinson-based first-principles distillation solver with a compact multilayer perceptron that learns condition-dependent corrections to binary interaction parameters involving the C 6 + pseudo-component. The equation of state and column solver are reformulated as a fully differentiable program in JAX, allowing end-to-end gradient propagation from plant-measured outputs to network parameters. Trained on 24 days of plant data, the hybrid model achieves mean absolute percentage errors below 0.3% for all monitored outputs, reducing mean absolute error by 94–98% compared to the baseline first-principles model. Leave-one-out cross-validation confirms generalization, with all outputs below 1% MAPE. The learned corrections are physically interpretable, and the prediction residuals identify operation outside the trained range, indicating the model's range of validity.