A transformer learns subsurface heat from sparse borehole data

Continental geothermal mapping relies on sparse, expensive borehole data, but conventional interpolation and physics-based models struggle with sharp thermal anomalies. In-Context Earth, a transformer, uses sparse local observations as context to predict continuous temperature-at-depth fields across the contiguous United States with 4.7°C mean absolute error—beating the physics-informed Stanford Thermal Model, AlphaEarth embeddings, and kriging. Without retraining, the model generalizes to Alberta, Australia, and the UK using only 20 observations at test time, achieving 2.2–6.2°C error. Uncertainty estimates are well calibrated, and interpretability analysis reveals the model learns unobserved subsurface properties (seismic velocity, geochemistry, crustal structure) and applies them consistently. Code and model weights are not mentioned in the abstract.