Prediction of electrode microstructure evolutions with physically constrained unsupervised image-to-image translation networks

Microstructure of electrodes determines the performance of electrochemical devices such as fuel cells and batteries. The efficiency and economic feasibility of these technologies depend on the stability of the microstructures throughout their lifetime. Although modeling techniques were proposed for determining electrode performance from 2- or 3-dimensional microstructural data, it is still extremely challenging to predict long-term structural degradation by means of numerical simulations. One of the major challenges is to overcome the difficulties in obtaining experimental data of an identical sample through the degradation process. In this work, a machine learning-based framework for predicting microstructural evolutions with limited amount of un-paired training data is proposed. Physically-constrained unsupervised image-to-image translation (UNIT) network is incorporated to predict nickel oxide reduction process in solid oxide fuel cell anode. The proposed framework is firstly validated by simplified toy-problems. Secondly, the UNIT network is applied to real microstructures of solid oxide fuel cells, which results in excellent visual and statistical agreements between real and artificially reduced samples. The proposed network can predict evolutions in new microstructures, which have not been used during training. Furthermore, a conditional UNIT network (C-UNIT) was demonstrated, which can predict the microstructure evolutions based on process conditions as well as continuous time series of microstructural changes.