Physics-Informed Neural Network with NSGA II and Levenberg–Marquardt Method for Kinetic Modeling in Heavy Oil Hydrocracking

Accurately modeling reaction kinetics in heavy oil hydrocracking is essential for optimizing reactor performance and improving product distribution predictions. This study proposes a novel hybrid framework that integrates a physics-informed neural network (PINN) with nondominated sorting genetic algorithm II (NSGA-II) and Levenberg–Marquardt (LM) optimization method to achieve fast and accurate estimation of kinetic parameters. Unlike conventional approaches, the proposed method combines global and local search: NSGA-II generates high-quality initial parameter estimates, while LM efficiently refines them, ensuring convergence within 300 epochs. This hybrid framework leverages a neural network to model time-evolving behavior, while a Runge–Kutta–based solver enforces reaction kinetics, enabling robust kinetic parameter estimation under physical constraints. Four kinetic models previously proposed in hydrocracking research were implemented as physics constraints and systematically evaluated using the PINN framework. Among them, the most detailed, referred to here as model 4, emerged as the most comprehensive and accurate, capturing all major saturates, aromatics, resins, and asphaltenes (SARA) conversions and byproduct formation (gas and coke). Building upon this, a refined 10-parameter kinetic model was proposed by excluding three low-sensitivity reaction parameters. The simplified model preserved all dominant pathways and demonstrated excellent predictive accuracy across four temperatures (360–400 °C), total error (data + physics MSE on mass fractions) was on the order of 10–3 to 10–2 across training, validation, and testing, with R2 between 0.93 and 0.99. To prevent overfitting and improve generalization, early stopping and a 20% dropout strategy were employed. This study presents a novel application of a hybrid PINN framework that integrates a multiobjective evolutionary algorithm with numerical optimization for kinetic modeling in heavy oil hydrocracking. By embedding physical constraints into the learning process, the framework offers a scalable, interpretable, and accurate approach for estimating reaction parameters and capturing the dynamic behavior of the hydrocracking process.