Optimizing vaccine supply chains under disruption risks using a neural network-enhanced variational inequality

This study develops a decentralized network-equilibrium model of the global vaccine cold chain that explicitly couples stochastic flood intensity with temperature-driven potency loss and embeds layered public-private risk-sharing instruments such as subsidies, public pools, and bilateral indemnities. The model captures the vulnerabilities and strategic interactions across manufacturing, distribution, and administration stages, using a variational-inequality framework that is solved by an extra-gradient algorithm warm-started by a shallow neural-network surrogate of the hazard-loss mapping, to achieve computational scalability. Numerical experiments show that an unmanaged flood cuts the vaccine flow in half and raises spoilage to 16.4 %. An optimized policy mix restores throughput by approximately 36.3 %, reduces spoilage by 56.7 %, narrows the shadow-price gap between upstream and downstream tiers by 68.4 %, and limits net profit loss per firm to no more than 4 %. The neural-network-accelerated solver reduces computation time and operator evaluations by approximately 50 % compared to the classical solver while preserving solution accuracy. These results demonstrate that relatively modest fiscal outlays can neutralize most disaster-induced performance losses. The analysis further shows that vaccine recipients’ willingness to pay critically determines equilibrium outcomes, while proactive mitigation strategies effectively align private and social costs. The framework offers both a computational advance and new policy insights for designing resilient vaccine logistics under climate risk.