Benchmarking for Scientific Machine Learning in Subsurface Geoscience

Our workshop aims to bridge the gap between scientific machine learning (SciML) and real-world geoscientific applications by developing and evaluating realistic, openly accessible benchmarks. Although SciML shows great promise, most methods are still only tested on simplified toy problems, holding back their relevance for complex systems such as subsurface heat transport, earthquake prediction, and CO₂ storage, as well as applications beyond geoscience.

Our event brings together method developers and geoscience experts from academia and industry to jointly define and implement stepwise benchmarks that evolve from academic setups to real-world case studies. These benchmarks are designed to connect the current state of SciML research with the practical requirements of geoscience modeling – bridging what is feasible today with what is needed in practice. They will address not only predictive accuracy but also robustness, efficiency, and usability. Interactive, hands-on sessions will allow participants to test and compare methods, identify limitations and future potential, define and refine benchmark designs. All developed benchmarks – including their definitions, datasets, reference solutions, and metrics – will be published to promote open collaboration and reproducibility.

By fostering interdisciplinary dialogue, the workshop aims to establish shared criteria for realistic SciML benchmarks, promoting transparency and reproducibility while paving the way for their broader adoption in geoscience. This collaborative effort will lay the groundwork for benchmark-driven innovation in SciML and strengthen the connection between SciML research and practical geoscientific challenges.

GITHUB link for the workshop: https://github.com/SCaLA-Group/SciML-GeoBench/