Climate-Informed Predictive Optimal Control of Desert Locust Dynamics Under Machine-Learning Climate Forcing.

Climate variability plays a fundamental role in desert locust population dynamics, yet most existing modelling and optimal control frameworks rely on simplified representations of seasonal forcing that inadequately capture short-term climatic variability. To address this limitation, we develop a climate-informed predictive optimal control framework that integrates a mechanistic stage- and phase-structured desert locust model with machine-learning-based climate forecasting. Long short-term memory networks and gradient-boosting regression are employed to predict temperature and rainfall, respectively, providing data-driven climatic inputs to a non-autonomous system describing locust development, reproduction, mortality, vegetation dynamics, and phase transitions. The mathematical analysis establishes the well-posedness of the model through the existence, uniqueness, positivity, and boundedness of solutions, while an autonomous reduction yields a closed-form basic offspring number that provides analytical insight into invasion potential under representative climatic conditions. A predictive optimal control problem is formulated to evaluate stage-specific physical, biological, and chemical interventions targeting juvenile and adult locust populations. Numerical simulations demonstrate that machine-learning-derived climate forcing more accurately reproduces observed climatic variability than conventional harmonic forcing, thereby providing improved inputs for predictive population modelling. Integrated juvenile-adult intervention strategies consistently outperform single-stage controls by reducing locust abundance, preserving vegetation, and lowering the composite management index. Among the intervention strategies considered, chemical control provides the greatest short-term suppression, biological control offers a more environmentally sustainable alternative, and physical control is most effective as a complementary measure during the early stages of population growth. Robustness analyses under deterministic climate perturbations and stochastic forecast errors show that the comparative ranking of intervention strategies remains stable under realistic climate forecast uncertainty. These findings demonstrate that integrating machine-learning climate prediction with predictive optimal control provides a mathematically rigorous and flexible framework for climate-informed decision support in desert locust management and establishes a foundation for future developments incorporating spatial dynamics, probabilistic forecasting, and real-time surveillance.