Neural Networks Predicting Submesoscale Tracer Dispersion

Abstract In this paper we examine the possibility of directly predicting passive tracer dynamics from flow fields. We use a two‐vertical‐mode model for generating flows ranging from low to high Rossby numbers and advect the passive tracer with such flows. While typically tracer dynamics are obtained by time integrating a tracer advection equation, here the emphasis is on predicting the tracer dynamics from the flow field. We experiment with popular architectures such as Autoencoder, UNet, and GAN, and also develop a novel model, that we call LoConv, to make predictions. The LoConv model uses custom Local Convolution layers that allows convolution with spatially varying weights and this model outperforms usually used architectures such as Autoencoder, UNet, and GAN based on various metrics. Autoencoders with very few trainable parameters were unsuccessful in making good predictions even at large‐scales. GAN and UNet predictions were biased towards large‐scale features, unfavorable for capturing small‐scale tracer dispersion, especially at high Rossby numbers. Overall, the LoConv model with some physics‐informed training produced the best fine‐scale tracer predictions, along with tracer flux and related derived quantities. More broadly, the results of this study point towards successful direct ways of predicting tracer fields from the flow, overcoming the computational cost of long numerical integration of tracer advection equations.