From Weather to Weathering: Foundation Models for Carbon Removal via Rock Weathering

Sasankh Munukutla, Felipe Oviedo, Kenji Takeda

Keywords: Foundational Models, Enhanced Rock Weathering, Carbon Dioxide Removal (CDR), Aurora, AI for Science, climate, interpretability

TL;DR: We show that latent representations from the weather foundation model Aurora transfer to predict rock weathering carbon dioxide removal rates, outperforming raw climate data across architectures.

Abstract: Enhanced Rock Weathering (ERW) harnesses natural silicate weathering, accelerating it by applying crushed rock powder onto soils. ERW has the potential to remove gigatons of carbon dioxide per year, making it a promising solution to stabilize Earth's climate. One of ERW's main challenges to being deployed at scale is cost-effective, rigorous, and accessible Measurement, Reporting, and Verification (MRV), which today requires dense ground-truth sampling and extensive laboratory analysis. We present TerraNova, to our knowledge the first deep-learning-based model that leverages latent weather representations learned by a Foundation Model (FM), Aurora, to predict Carbon Dioxide Removal (CDR) rates of rock weathering in real-world field conditions. Trained on over 170,000 chronosequences of largely natural weathering data and limited ERW data, TerraNova achieves $R^2 = 0.8261$ and MAE $= 0.0750$ on held-out validation sites, demonstrating strong performance across diverse conditions. Through interpretability analysis on synthetic ERW scenarios, we demonstrate alignment with existing geochemical understanding across several key weathering drivers, e.g., higher soil moisture leads to higher CDR rates, showing that TerraNova learns the underlying drivers from largely natural weathering. Our results show the promise of AI and FMs to predict carbon removal and the potential to aid ERW MRV at climate-relevant scales.

Submission Number: 124