Advancing AI for Earth sciences with hybrid and causal models

Understanding and predicting Earth system processes requires more than accurate forecasts—it demands uncovering the underlying relationships among variables and constructing models that are interpretable, physically consistent, and mathematically robust. While machine learning has demonstrated exceptional predictive capabilities, its models often neglect fundamental physical laws, raising concerns about reliability, interpretability, and trust. This work explores the integration of domain knowledge with machine learning through hybrid and causal modeling approaches, aiming to bridge data-driven insights with the principles of the physical sciences. By leveraging these methods, we can enhance our understanding of the data-generating processes and achieve results that are both consistent and explainable. I will present recent advances and strategies in this field, highlighting their potential to revolutionize Earth system research. This effort represents a step toward a long-term AI agenda for developing algorithms that drive knowledge discovery in Earth sciences.

 

https://arxiv.org/pdf/2010.09031.pdf

https://arxiv.org/abs/2104.05107.pdf

https://doi.org/10.1016/j.physrep.2023.10.005