Modelling of solid electrolyte interphase growth using neural ordinary differential equations

In this work, neural ordinary differential equations (NODE) are used to identify phenomenological growth rate functions to model the solid electrolyte interphase (SEI) growth during formation. To analyse the capabilities of this approach in a controlled setting, synthetic SEI thickness data is generated using a model that uses a mechanistic growth rate function. Several possible implementations and extensions of the NODE are investigated, including physical constraints and data augmentation. All the investigated variants agree well with the training data, but significant differences are observed for the validation data. The results show that the growth rate functions learnt by the baseline implementation without further constraints significantly differs from the growth rate function given by the mechanistic model. However, it is shown that the use of appropriate data augmentation or physical constraints provides a significant improvement, and low errors can be achieved within the validation data sets. It is concluded that NODE can reveal growth rate functions, but careful consideration is needed to achieve functions that are phenomenologically consistent with the underlying mechanisms.