A Machine Learning‐Based Geothermal Gradient Framework for Constraining Lithospheric Biomass

Abstract

Earth's deep subsurface hosts a large microbial biosphere, but the magnitude of lithospheric biomass remains poorly constrained due to sparse observations and uncertain thermal limits. Here, we present a physically constrained global assessment of the continental and oceanic lithospheric biomass shallower than the 122°C isotherm. Using 8,452 geothermal gradient observations aggregated to a 1° × 1° grid, we trained separate XGBoost models for continental and oceanic domains to reconstruct a continuous global geothermal gradient field. Temperature‐bound habitable depths derived from this framework yield comparable habitable volumes of 6.8 × 10 ⁸  km ³ for the continental lithosphere and 7.3 × 10 ⁸  km ³ for the oceanic lithosphere. Recalculated continental biomass agrees with previous estimates, yielding 2–6 × 10 ²⁹ cells (4.2–12.6 Gt C). In contrast, oceanic biomass is highly sensitive to the treatment of shallow seawater‐influenced samples across alternative extrapolation schemes. Excluding these samples yields estimates of 0.5–1.6 × 10 ²⁸ cells (0.1–0.3 Gt C), substantially lower than previous estimates, whereas including them inflates biomass estimates by up to 4 orders of magnitude. Combined continental and oceanic biomass estimates total 4.3–12.9 Gt C. These results indicate that lithospheric biomass is more limited than previously inferred and demonstrate that physically constrained integration provides a robust framework for quantifying the deep biosphere.