A machine learning framework for modeling long-term forest carbon dynamics under warming and nitrogen enrichment

Forest carbon could be used as a substantial global sink for atmospheric carbon through active forest management practices, helping forest productivity and reducing atmospheric concentration of greenhouse gases. However, predicting long-term forest carbon responses to concurrent warming and nitrogen (N) enrichment remains a critical challenge for Earth system science. Although process-based biogeochemical models such as ForCent provide mechanistic insights on soil organic matter (SOM), they are computationally expensive and require extensive parameterization. Here, we develop a novel machine-learning (ML) framework that integrates temporal feature engineering with multi-task Bayesian regression (MTBR) to emulate and accelerate ForCent simulations for the Harvard Forest red oak ecosystem. The ML models reproduced SOM pool dynamics with R 2 ≈ 0.99 and predicted growing-season carbon accumulation with R 2 ≈ 0.97, while reducing error by up to 90% and computational time by over an order of magnitude. Feature attribution identified trend-lagged above-ground carbon and soil temperature as dominant controls, consistent with known ecological drivers. The proposed ML recovered canonical treatment responses for SOM decline under warming and biomass enhancement under N addition, closely aligning with ForCent and field observations. This study demonstrates that ecological–ML modeling can preserve process fidelity, offering a scalable pathway for next-generation forest carbon forecasting under global environmental change. • Machine-learning emulates ForCent carbon–nitrogen process models. • Captures long-term forest carbon dynamics with ≈ 0.99 R2 accuracy. • Run time reduced by tenfold relative to process-based simulations. • SHAP analysis reveals key carbon and temperature interactions. • Framework supports carbon-neutral forest management decisions.