Incorporating Physical Constraints inside Neural Networks to Improve their Accuracy and Physical Reliability for Chemical Engineering Unit Operations Modeling

Abstract

Neural networks are machine learning models structured in interconnected layers of nodes, or neurons, designed to process and learn complex data patterns by adjusting connections based on the input data and desired output. A common challenge with these networks lies in their limited ability to incorporate fundamental physical principles, as they typically prioritize data pattern recognition over adherence to system-specific laws and constraints. This research introduces an advanced modeling framework that integrates physics-informed neural networks with data reconciliation techniques, embedding physical constraints directly into the neural network's learning process. By enforcing consistency with foundational physical laws, this hybrid approach effectively combines data-driven insights with physics-based accuracy, enhancing the model’s reliability for complex engineering applications. The study further assesses the performance of traditional neural networks, physics-informed networks, and data reconciliation methods, focusing on their application in the design and optimization of unit operations, revealing the advantages of this physics-augmented approach in bridging theoretical principles and practical modeling.