Accurate Near Fields Prediction from Nanosphere with Physics Driven Deep Learning

This paper proposes a physics-driven deep learning model, which defines the custom loss function based on the wave equation. The average relative norm error of the test set is 3.88%, while it would be 6.36% with standard MSE loss function. Compared with training only using MSE as loss function, the accuracy of the physically driven model is enhanced, and it improves the convergence rate, especially at early training, as well as the electric field figure quality.