AI-Driven Mechanistic Modeling of Biological Processes for Drought-Resilient Crop Design

Abstract Living systems respond to environmental constraints through multiscale interactions between genetic potential and physiological processes. Mechanistic modelling encodes this causality, and its strengths in calibration and prediction can be extended by workflows that interrogate model structure to support transfer across environments. This study contributes an integrated diagnosis–design framework that unifies: (i) satellite-driven characterization of environmental structure (terrain–soil–climate), (ii) mechanistic–data-driven benchmarking to quantify residual error after calibration, and (iii) bounded inverse engineering that explores phenotype space under physiological constraints and projects computational optima onto field-characterized cultivars. Using rainfed aerobic rice, 3D PCA–Gaussian mixture clustering identifies twelve environments from soil–climate attributes. Despite optimized calibration, CERES-Rice plateaus ( r ² = 0.049–0.200 ) and predicts water-use efficiency (WUE) magnitudes ( max ~0.24 kg ha⁻¹ mm⁻¹ ) inconsistent with drought-adapted aerobic ranges. An ARX/NARX benchmark reaches r ² = 0.99 for grain yield and r ² = 0.80 for biomass/grain number, recovering WUE in the aerobic range ( ~2–15 kg ha⁻¹ mm⁻¹ ) while remaining mechanistically opaque. This motivates probing mechanistic behaviour across the constraint space. Inverse engineering quantifies a 22–30% gap to optima and maps them onto a panel of 21 field-validated cultivars, highlighting WAB56-50 and DKAP2 as closest candidates.