A Generalized Framework for Machine Re-learning of Complex Process and Kinetic Models

Abstract

Computer modelling is commonly used to simulate complex systems and thus give insight to the problems with lower cost and more efficiently. However, since they usually characterize real system by using complicated mathematical equations, it could be very time-consuming and expensive to run them on the computer. To tackle this problem, artificial neural network (ANN) model, in many cases, has turned out to be successful surrogate model to reduce the computational burden and generate data in faster manner. However, though there is plenty of software available to create neural network, it still takes time to determine every detail for a good model. Therefore, this study was motivated to build a Python module which builds a multi-layer perceptron (MLP) neural network and searches the best model architecture to substitute a numerical simulation model. A synthetic data has been utilized to prototype the module. After every function works out with synthetic data, data from Pseudo-two-dimensional (P2D) battery model has been employed to test the functinality of the module.However, the best prediction gained from over 2000 models trained gives R2 as only 0.3245. There are several possible causes of not finding a useable model, which could be improper processing of data, unsuitable choices of activation functions, not enough iterations went through,inappropriate optimization method, etc. The module thus needs more work to be perfected.