Mechanistic agroforestry simulators, such as Hi-sAFe, provide faithful representations of trees×crops interactions, but impose computational costs that preclude large-scale parameter space exploration. We present MetAIsAFe: a complete machine learning meta-modelling pipeline that replaces Hi-sAFe simulations with near-instantaneous predictions while preserving physical coherence over 40-year trajectories. An input space of approximately 20 parameters was sampled following Sobol and LHS strategies to generate several campaigns of unitary simulations, spanning geography, soil texture, plot geometry, crop species, and climate scenario, each simulated in three variants (agroforestry, sole-crop, sole-forest) under the RCP-8.5 climate scenario. A first sensitivity analysis campaign of 2048 Sobol configurations (sobol_S11111_n2048) informed a refined training campaign of 2044 Sobol configurations (sobol_training_1_n2048). A second training campaign of 3169 LHS simulations (lhs_training_2) was generated for more precise meta-model training. A cascade architecture comprising two binary classifiers (tree viability, crop yield failure), ten discrete-horizon regression models for tree stem carbon (LightGBM + PCHIP interpolation), and two row-by-row yield regressors was designed. The pipeline achieves R2 = 0.817 (ρ = 0.930) for agroforestry stem carbon at year 40, R2 = 0.686 for agroforestry crop yields, R2 = 0.645 (ρ = 0.800) for sole-forest stem carbon at year 40, and R2 = 0.619 for sole-crop yield predictions on held-out test sets. The binary classifying pre-filter rejected 67% of 10 000 candidate configurations prior to simulation for the second training batch, saving an estimated 55 000 hours of cluster time. Global temporal sensitivity analysis using the normalised Hilbert-Schmidt Independence Criterion (HSIC) with bootstrap confidence intervals revealed a marked temporal transition: soil texture parameters (sand, clay, stone) dominate crop yield variance across all 40 years, while stem carbon sensitivity is governed by clay content during the juvenile establishment phase (years 1_15) before transitioning to plot geometry and soil depth in the mature phase. The climate period (PRE vs. FUT) exerts a detectable but secondary effect on sole-crop yields (HSIC = 0.025) but is nearly neutral for carbon targets. MetAIsAFe provides a complete ML meta-modelling pipeline for Hi-sAFe, reducing 40-year trajectory prediction time from several hours to under 100 ms, and enabling the first temporal HSIC sensitivity analysis of a detailed agroforestry simulator.
Publication Date: 2026-06-18