A comparative study of deep learning models for two-dimensional MXene ferroelectric data

Recently, two-dimensional MXenes have gained attention for use in electronics, sensing, and energy storage as a potential family of layered materials with adjustable ferroelectric behavior. In this work, we were investigating whether deep learning models could speed up ferroelectric analysis by predicting the time dependence of MXene-based ferroelectric properties by using previous values only. We chose an autoregressive univariate model, in which each of the properties Capacitance, Dielectric Constant, Instantaneous Current, and Measured Polarization were modeled by their past behavior. We benchmarked six neural architectures with the same training conditions given a fixed set of hyperparameters: Long short-term memory, gated recurrent unit, bidirectional long short-term memory, convolutional neural network–long short-term memory, convolutional neural network–gated recurrent unit and convolutional neural network–bidirectional long short-term memory, to provide comparability. Root mean squared error and mean absolute error measures of model performance were assessed on held-out data. Findings indicated that CNN+GRU obtained the least error in Instantaneous Current whereas GRU was the most suitable in Measured Polarization, Capacitance and Dielectric Constant. It is interesting to note that the most efficient architectures to train, including CNN+LSTM and Bi-LSTM, did not necessarily perform well on generalization, an observation that training loss is not a good model selection measure. This work demonstrates that various MXene ferroelectric characteristics necessitate different deep learning architectures, and the modeling of the specific properties demands property-specific models. • Deep learning predicts MXene ferroelectric behavior from historical data. • Evaluation of six neural models using the same hyperparameters. • GRU best for other properties; CNN-GRU best for Instantaneous Current. • Property-specific architecture outperforms universal models. • Compared to training loss, validation loss is a more accurate measure of generalization.