PINN-XAF-MTL: A Multi-Modal Deep Learning Framework Integrating Novel Physics-Informed Cross-Attention Fusion for Predicting Multi-Environment Hybrid Performance

Abstract

Artificial intelligence (AI) is redefining hybrid breeding by shifting the operational paradigm from empirical field screening to predictive representation learning. While next-generation sequencing and high-throughput phenomics have generated massive multi-omics datasets, classical linear models fail to resolve the non-linear epistatic networks, dominance variances, and dynamic Genotype × Environment (G×E) interactions that govern heterosis and combining ability. This review provides a novel, critical synthesis of AI architectures in hybrid prediction, categorizing deep learning paradigms based on their biological inductive biases. We dissect the transition from single-modality models to intermediate cross-attention multimodal fusion networks that synthesize genomic markers, latent-space phenomics, and temporal environmental covariates. Furthermore, we address critical field bottlenecks—specifically the "Diallel Sparsity Paradox," population structure confounding, data privacy barriers, and the out-of-distribution generalization failures of models facing climate anomalies. To resolve these, we propose the integration of Explainable AI (XAI) for biological validation, Federated Learning frameworks for cross-institutional collaboration, and Physics-Informed Neural Networks (PINNs) that embed deterministic crop physiological equations into deep learning loss functions. Finally, we map out an operational roadmap for Executable Digital Twin (xDT) breeding ecosystems, establishing a definitive blueprint for next-generation, climate-resilient hybrid development.