Hybrid neural networks for improved chemical process modeling: Bridging data-driven insights with physical consistency
Year: 2025
Type: article
Abstract: The increasing reliance on neural networks (NN) in chemical process modeling highlights their capability for accurate predictions, yet their standalone application often struggles to adhere to fundamental physical laws such as equilibrium constraints and mass balance. Addressing this limitation, hybrid methods that integrate data-driven insights with physical consistency have gained prominence. This study investigates the integration of NNs with nonlinear data reconciliation (NDR), proposing it as a robust approach for enhancing the reliability of chemical process models. Using a Gibbs reactor as a case study and extending to reactor-distillation integration, we evaluate hybrid methodologies including NN x NDR, NN x KKT (Karush-Kuhn-Tucker), and KKT-PINN (Physics-Informed Neural Networks with KKT conditions). The proposed method of merging NNs with NDRs demonstrates superior performance in minimizing errors and enforcing physical constraints. Its robustness is validated under various conditions, showcasing scalability and applicability to multi-unit systems. This work underscores the transformative potential of hybrid approaches in achieving both data-driven accuracy and physical system fidelity.
Source: Digital Chemical Engineering
Authors: Jana Mousa, Stéphane Negny, Rachid Ouaret
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,650
OpenAlex ID: W4413271014