Development of mass, energy, and thermodynamics constrained steady-state and dynamic neural networks for interconnected chemical systems

Abstract

This paper discusses the development of steady-state and dynamic modeling algorithms for mass, energy, and thermodynamics constrained neural networks (METCNNs) for interconnected chemical process systems. The METCNN models can ‘exactly’ conserve the overall system mass and energy balances, as well as certain thermodynamics constraints during both training and forward problems. The proposed approaches can accommodate an outer layer integer programming problem for selection of the best thermodynamics model from a family of candidates given a particular transient dataset. The developed algorithms for both steady-state and dynamic METCNNs are tested for an interconnected chemical system in presence of noise and bias in training data. For all case studies considered in this work, it has been observed that the optimal METCNN models ensure exact conservation of system physics and consistently converge close to the system truth, even when trained against complex dynamic noisy measurements that do not necessarily satisfy the system physics.