Neural Network Programming: Integrating First Principles into Machine Learning Models

Abstract

Tired of borrowing the machine learning models proposed by computer scientists? This thesis proposes a new perspective, ideas, and methods to formulate hybrid models through the manipulation of the Artificial Neural Network architectures and parameters. It is shown that the actual transcription of exact physics into a neural network is possible. This allows leveraging all the advantages of neural network modelling, namely, universal approximation feature, autodifferentiation, efficient model evaluation, and advanced optimization schemes. It might be possible to convince even the skeptics that Machine Learning is a feasible tool for the digitalization of chemical engineering processes. No longer we should be constrained by the tyranny of the typical architectures of the Neural Networks!