Exploring the Response of Maximum Carboxylation Rate to Drivers Through an Interpretable Machine Learning Framework

The maximum carboxylation rate (Vcmax) is a key parameter determining photosynthetic capacity in terrestrial biosphere models, yet its variability is often oversimplified as a plant functional type-dependent constant, increasing uncertainties in gross primary productivity (GPP) simulations. In this study, we inverted the daily Vcmax normalized to 25 °C (Vm25) by coupling the Breathing Earth System Simulator (BESS) model with the light response curves (LRC) method using eddy covariance observations at five maize sites in the United States and China. Then, four machine learning (ML) models—KNN, SVM, Random Forest (RF) and XGBoost—were employed to simulate Vm25, with RF showing the best performance (R2 = 0.83, RMSE = 7.57 μmol m−2s−1 for testing). Cross-validation further demonstrated the RF model’s ability to capture seasonal trends across sites. Incorporating the RF-simulated Vm25 into the BESS model significantly improved GPP estimates compared to the original BESS, with R2 increasing from 0.53–0.81 to 0.77–0.84. Through the SHAP method and ablation experiments, leaf age was identified as the most influential factor, with the largest SHAP value of 8.37 μmol m−2s−1, higher than that for temperature (4.47 μmol m−2s−1), solar radiation (4.21 μmol m−2s−1) and leaf area index (2.32 μmol m−2s−1). And vapor pressure deficit had the minimal SHAP value of 0.87 μmol m−2s−1. Notably, the effect of leaf age on Vm25 exhibited a unimodal pattern, with a strong coupling effect with leaf area index. This study demonstrates that interpretable machine learning not only provides a robust approach for simulating seasonally dynamic Vcmax, but also enhances our understanding of its driving biological and environmental factors, offering a valuable pathway for improving carbon cycle modeling in agroecosystems.