Peeking Into a Marine Biogeochemical Model with an Autoencoder

Biogeochemical models are computational approximations of systems of differential equations used to represent and predict biogeochemical constituents of the ocean. These might include carbon and nutrients cycles, and its interactions with biological components, such as different types of plankton. Unfortunately, these models tend to be constrained by their complex parametrization and computational cost, limiting their practical application and scalability.

Autoencoders are neural networks that are trained to learn a compressed representation of a dataset, typically with the goal of reconstructing the input to its original or a specified target dimension. But the bottleneck of this compression, or the latent space of the autoencoder, can offer interesting insights into the dominant features of the system.

Here we train different types of autoencoders to capture the main spatiotemporal dynamics from data modeled by the biogeochemical analysis BIO4 (based on NEMO-PISCES). This not only provides a basis for the development of computationally efficient emulators, but it can help us detect patterns and relationships that might not be immediately apparent in the high-dimensional output of the model. This offers interesting insights into how the model actually captures its constituting components. 

Such compressed representations can also be used for parameter sensitivity analysis, to develop data assimilation frameworks, and as tools for uncertainty quantification and outlier detection.