Neural network programming: Integrating first principles into machine learning models
Andrés Carranza-Abaíd, Jana P. Jakobsen
Abstract
This work introduces Neural Network Programming (NNP) as an integrated hybrid modelling approach. NNP consists in formulating a set of first principles equations that is later decomposed and transcribed into an Algorithmically Structured artificial Neural Network (ASNN). NNP leverages the advantages of the universal approximation theorem and neural network optimization algorithms in order to generate physically coherent machine learning models. Since ASNNs are not mere approximations of physics equations, it is not necessary to modify either the gradient or performance function in order to account for errors with respect to the first principles equations. ASNNs are trained faster and more accurately than typical hybrid models because the gradient is computed through automatic differentiation instead of numeric differentiation. It is shown that the same ASNN architecture is transferable between processes with similar characteristics. In particular, a flash separator, distillation column, and a biogas upgrading process, were modelled using an identical architecture.