Foundation models for predicting enhanced rock weathering at operational scale

Enhanced rock weathering (ERW) is a promising carbon removal solution, but quantifying removal requires extensive physical sampling. We present TerraNova, a deep learning model that combines pretrained weather representations with encoded soil and rock-amendment data to predict near-field carbon removal. We select its architecture on 192,511 natural-weathering measurements from 185 locations, then train and evaluate it on 5,192 ERW measurements from 264 fields in the United States and Brazil. In grouped cross-validation across 64 sites, TerraNova predicts carbon removal with at least twice the predictive skill of baseline models, and its accuracy degrades least as sites lie farther from the training data. Without on-site measurements, it identifies five of the ten highest-removal sites among 52 candidates, versus at most two for baselines. Combined with one measurement, it matches six measurements alone, an 83% sampling reduction at the median site, and across the seven largest sites it reduces sampling by 75% and 63% for target errors of 0.75 and 0.5 t CO$_2$e ha$^{-1}$ yr$^{-1}$.