Hard-Constrained Mass Conserving Neural Network Modelling for Industrial Manufacturing Processes
Please login to view abstract download link
In process manufacturing, especially in systems governed by physical laws like mass conservation, ensuring that machine learning models respect fundamental constraints is critical for model reliability. Traditional neural networks, while flexible and powerful, do not inherently adhere to conservation laws, which can result in unphysical or unrealistic predictions when applied to processes such as semi-batch reactors. To address this limitation, we develop a mass-conserving neural network (MCNN) framework that embeds hard constraints, by construction, directly into the model to guarantee mass conservation at every prediction step. This approach was applied to a semi-batch chemical process where only initial conditions and some process conditions are known. The model was trained on simulated process data while ensuring strict adherence to mass conservation. Results demonstrated that the MCNN accurately predicts the system’s states and inherently maintained mass balance, providing reliable and physically consist
