Synergizing Physics-Informed Transformers and Gradient Boosting for Resilient Crop Yield Prediction in Arid Environments: A Case Study in Oman

Agriculture in arid regions like Oman faces significant challenges due to water scarcity and extreme heat. This study introduces a hybrid physics-informed machine learning framework that integrates Gradient Boosting Decision Trees (GBDT) with Temporal Fusion Transformers (TFT). Unlike purely data-driven models, our approach embeds biophysical constraints—specifically the Water Use Efficiency (WUE) limit—into the neural network loss function. Optimized through Bayesian methods (λphys=1.965), the model demonstrates superior resilience by reducing prediction errors by 15% under extreme heat stress (>45°C) while ensuring 100% compliance with biological limits. Validated through a global data competition, the framework achieves top-tier predictive accuracy. This research provides a scalable decision-making tool for resource-efficient agriculture, supporting the food security goals of Oman Vision 2040. Keyword: Physics-Informed Machine Learning, Oman Agriculture, Crop Yield Prediction, Water Use Efficiency, Temporal Fusion Transformer.