Time-series neural networks to predict electromagnetic wave propagation

Understanding wave propagation is fundamental across numerous scientific domains, underpinning crucial tasks in acoustics, seismology, radar technology, materials science, and optics. Machine learning methods offer a promising avenue to deepen our understanding of wave propagation dynamics, providing insights into the behavior of near-field wave patterns. Moreover, well-trained machine learning models have the capacity to generalize beyond specific training data, allowing for predictions in scenarios not explicitly encountered during training. This paper presents a machine learning approach using time-series neural networks to predict the complex near-field wave patterns emerging from metasurface devices. The recurrent neural network (RNN) and the long short term memory (LSTM) models are presented along with a custom dataset that includes 3x3 configurations of meta-atoms. The experiment focuses on assessing the models' capabilities with varying amounts of input data and explores the challenges posed by predicting propagating waves. Results indicate that the LSTM outperforms the RNN, markedly in learning training data, highlighting its efficacy in capturing complex dependencies. Analysis of error metrics reveals insights into the impact of dataset size on model performance, with larger datasets posing computational challenges but potentially enhancing generalization. Overall, this study lays the foundation for advancing the use of time-series machine learning models for applications involving wave propagation, with implications for various applications in photonics and beyond.