Predicting productivity and climatic resilience of coniferous–broadleaved mixed plantations with 3-PGmix and machine learning

The transition from monoculture plantations to mixed-species stands is a key strategy in China for improving productivity and climatic resilience. However, the long-term productivity outcomes of different transformation measures, particularly under future climate change, remain insufficiently quantified. Here, we developed a hybrid modeling framework combining the process-based 3-PGmix model with machine learning to simulate and efficiently extrapolate spatiotemporal dynamics of biomass and net primary productivity (NPP) over 40 years following mix-species transformation. In this framework, machine learning serves as a computational surrogate to emulate 3-PGmix outputs, enabling rapid scenario exploration while retaining the structural assumptions of the underlying process model. Results indicate that the effects of species mixing on stand productivity are stage-dependent, with the strongest productivity gains occurring in mixed stands containing approximately 70% coniferous and 30% broadleaved species. In the study area, converting pure Cunninghamia lanceolata plantations to mixed stands with Schima superba at this proportion is projected to increase total productivity by 3336.74 t year −1 by 2060 (equivalent to 1.2 t ha −1 year −1 ). Across both SSP245 and SSP585 scenarios, mixed stands consistently exhibited lower relative biomass decline than pure C. lanceolata stands—particularly during mid-rotation stages—indicating greater stability under projected climate trajectories. Sensitivity analysis shows that productivity losses are primarily driven by temperatures exceeding optimal growth thresholds. Maximum temperature effects outweighed the compensatory influences of precipitation and minimum temperatures. For computational efficiency, Random Forest (RF) was selected as a surrogate emulator of 3-PGmix simulations. It outperformed artificial neural networks (ANN) and multiple linear regression (MLR) in reproducing simulated NPP patterns, demonstrating high approximation accuracy and numerical stability within the modeled parameter space. This hybrid framework therefore functions as a computationally efficient extension of 3-PGmix for large-scale scenario analysis. It provides an interpretable platform for evaluating plantation transformation strategies under projected climate change in subtropical China. • Species-mixing effects on plantation productivity are stage-dependent, with maximum gains occurring in stands composed of ~70% coniferous and 30% broadleaved species. • Transforming pure C. lanceolata plantations to mixed stands with S. superba is projected to increase total productivity by 3336.74 t yr −1 (1.2 t ha −1 yr −1 ) by 2060. • Mixed stands exhibit enhanced climate resilience under both SSP245 and SSP585 scenarios, showing consistently smaller relative biomass declines than pure coniferous stands. • A Random Forest surrogate model efficiently reproduces 3-PGmix-simulated spatial NPP patterns, outperforming ANN and MLR in accuracy, stability, and computational efficiency.