DJ4Earth is a cross-university collaborative project that aims to build open-source, differentiable Earth system models using Julia, with derivative codes enabled via the automatic-differentiation (AD) tool Enzyme. Three of the major components of the project are the implementation of AD-enabled large-scale sea and land ice models, as well enabling AD for an existing ocean general circulation model that is part of the Climate Modeling Alliance (CliMA). A key feature of these geophysical models will be the use of Enzyme for sensitivity analysis, data assimilation, scientific machine learning, and predictive modeling through the implementation of the so-called adjoint method (reverse mode AD). AD has been used for computing adjoints (i.e., gradients) in the past, though largely using propriety software. DJ4Earth aims to open the door for a much broader research community to view and work with these models.

In addition, the creation of these open-source projects will advance the checkpointing, parallelization support, and sparsity optimization capabilities within Enzyme. The goal of this roundtable will be to facilitate discussions regarding the needs for DJ4Earth projects and how Enzyme is incorporated within them, as well as how Enzyme can improve the capabilities of said projects; for example, Enzyme's AD capabilities may enable the use of concepts from scientific machine learning within our models, such as data-driven surrogates and parameter/state estimation. Overall, Enzyme brings the potential for more flexible, open-source automatic-differentiation capabilities in the Earth sciences.