Oral A3 Data-Driven Emulators

4:15 PM - 4:30 PM JST(7:15 AM - 7:30 AM UTC)

[hyhteQZr-02] Differentiable Earth: Theory and Practice

*Takeshi Enomoto¹ (1. Kyoto University (Japan))

Many problems in Earth science are inverse problems that require parameter, initial, or boundary condition estimation given a model and data. The adjoint method has traditionally provided an efficient solution to these problems. In atmospheric and ocean prediction, variational data assimilation (VDA) utilizes an adjoint model. However, the development and maintenance of an adjoint model require significant effort compared to the popular ensemble-based approach. The rise of machine learning (ML) has renewed interest in this method: since both ML and VDA share a mathematical foundation in optimization, VDA can now directly leverage modern ML frameworks that offer automatic differentiation (AD). Unlike traditional source-to-source translators, AD generates derivatives for optimization without explicit adjoint code compilation. We argue that AD functions similarly to the systematic approach of building an adjoint using the Lagrange multiplier method (Thacker and Long 1988). In this talk, we demonstrate AD-based VDA (ADVDA) applied to idealized and intermediate complexity models. We show that ADVDA replicates results from hand-coded adjoints with comparable efficiency, and we examine ADVDA performance in the presence of discontinuities. These findings demonstrate the utility of ADVDA in solving inverse problems—provided careful attention is paid to technical challenges such as computational efficiency, memory management and discontinuities—and motivate us to build differentiable digital twins.

Acknowledgements: This work was supported by JSPS KAKENHI Grant Number 24H02226 and JST K-Program Grant Number JPMJKP25Q1.

References: Thacker, W. C., and R. B. Long, 1988: Fitting dynamics to data. J. Geophys. Res. Oceans, 93, 1227–1240, doi:10.1029/JC093iC02p01227.