DIRESA – A Deep Learning-based, nonlinear "PCA"​

A deep ANN-based dimension reduction (DR) method, called DIRESA (distance-regularized Siamese twin autoencoder), has been developed to capture nonlinearities while preserving distance (ordering) and producing statistically independent latent components. The architecture is based on a Siamese twin autoencoder, with three loss functions: reconstruction, covariance, and distance loss. An annealing method is used to automate the otherwise time-consuming process of tuning the different weights of the loss function. DIRESA has been compared with PCA and state-of-the-art DR methods for two conceptual models, Lorenz ’63 and MAOOAM (Modular Arbitrary-Order Ocean-Atmosphere Model), and significantly outperforms them in terms of distance (ordering) preservation KPIs and reconstruction fidelity. The latent components have a physical meaning as the dominant modes of variability in the system. DIRESA correctly identifies the major coupled modes associated with the low-frequency variability of the coupled ocean-atmosphere system. Next to the conceptual model results, the first DIRESA results for reanalysis data will be presented.

DIRESA is provided as an open-source Python package, based on Tensorflow. With one line of code convolutional and/or dense layers DIRESA models can be build. On top of that, the package allows the use of custom encoder and decoder submodels to build a DIRESA model. The DIRESA package acts as a meta-model, which can use submodels with various kinds of layers, such as attention layers, and more complicated designs, such as graph neural networks. Thanks to its extensible design, the DIRESA framework can handle more complex data types, such as three-dimensional, graph, or unstructured data. Its flexibility and robust performance make DIRESA an promising new tool in weather and climate science to distil meaningful low-dimensional representations from the ever-increasing volumes of high-resolution climate data, for applications ranging from analog retrieval to attribution studies.

Tutorial: https://diresa-learn.readthedocs.io/

Preprint: https://arxiv.org/abs/2404.18314