Generative adaptive resilience: LLM-MILP modeling and strategy generation for port vessel scheduling

Port scheduling systems face critical challenges from escalating disruptions and the limitations of traditional optimization, including rigid feasibility checks, opaque decision-making, and the absence of learning mechanisms. This paper introduces the generative adaptive resilience (GAR) framework, an integrated approach that advances port scheduling from static constraint satisfaction toward a dynamic, learning-enabled process. The GAR integrates generative AI (GAI), mathematical optimization, and evolutionary learning into a closed loop, enabling ports to respond to disruptions and proactively prepare for them. The framework bridges large language models (LLMs) with operations research through the Feasibility-Aware Constrained Decoding (FACD), converting human-interpretable actions (e.g., redirect, delay, resequencing) into mathematically feasible scheduling adjustments with a feasibility guarantee. The GAR explicitly models and exploits synergistic action pairs (e.g., jointly delaying a feeder vessel while redirecting a deep-sea liner), thereby achieving cost reductions through coordination coefficients. Empirical validation using Ningbo Port data shows that the GAR achieves faster policy convergence, reduced high-priority vessel delays, and faster solution time than full-resilience baselines, while maintaining equivalent redundancy. The framework maintains safe collaborative action rates and reduces peak terminal utilization. These results demonstrate the GAR’s concrete advantages: accelerated adaptation, quantifiable service improvements, and computational efficiency, providing ports with a measurable pathway from reactive operations to adaptive, learning-enabled resilience, contributing to risk and safety science in the AI Era. • A framework of GAI and a mathematical program for adaptive port resilience. • Constrained decoding enables the Large Language Model to create schedule actions. • Explicitly optimize peak utilization to prevent terminal overload and delay. • Validated on Ningbo Port data: faster convergence, lower delays, higher synergy. • Enable multi-port and privacy-preserving transportation resilience evolution.