Integrating digitally enhanced data extraction and simulation modelling for AI-driven supply Chain resilience: an operational research framework for strategic stockpiling of critical minerals

Simulation modelling in Operations Research (OR) relies critically on data quality, yet multi-echelon supply chains (SC) often lack the timely, contextual information necessary for realistic model parameterisation and scenario generation. This research proposes the “Large Language Models (LLM)-Integrated Simulation Data Framework”, a domain-agnostic methodology that systematically integrates unstructured narratives with structured datasets to enhance simulation model realism and contextual relevance. The framework employs a Retrieval-Augmented Generation pipeline powered by LLMs to extract, structure, and validate intelligence from policy documents, news archives, and industry reports, transforming qualitative narratives into traceable model inputs for parameterisation, behavioural logic, and scenario specifications. We demonstrate the framework’s applicability through India’s strategic development of its lithium carbonate reserves processing 972 documents to generate 668 validated insights that inform agent behaviours, numerical parameters, and scenario specifications. The methodological contribution establishes a novel collaborative interface between OR and Generative Artificial Intelligence, demonstrating how automated extraction of contextual knowledge at scale enables the combination of quantitative baselines with qualitative intelligence to enhance simulation empirical grounding. The lithium stockpiling case validates the framework’s applicability in data-sparse, multi-echelon contexts, yielding policy insights on strategic reserve development whilst demonstrating broader applicability to geopolitically sensitive, resource-constrained SCs. For Operations Research (OR) practitioners and policymakers facing incomplete, fragmented, or rapidly changing data, this research introduces the “LLM-Integrated Simulation Data Framework”. This domain-agnostic methodology automatically extracts and structures intelligence from unstructured documents at scale, then integrates these insights with traditional databases to build realistic, evidence-based simulation models through informing model logic, input modelling, and scenario generation. This framework enables partitioners to develop simulation models for emerging risks (e.g., regulatory changes) when structured data are limited, outdated, or unavailable. It automates the extraction and organisation of information from extensive unstructured sources, reducing time-consuming manual review while maintaining full traceability for human validation. The framework is modular, domain-agnostic, and practical to deploy: although running large LLMs locally can require stronger hardware, this constraint can be reduced through lightweight models or API-based deployments. We demonstrate the framework through India’s strategic lithium carbonate reserve and derive several policy-relevant insights. Strategic reserve effectiveness depends on coordinated management of acquisition timing, sourcing strategy, and release protocols. Counterintuitively, concentrated sourcing outperforms geographic diversification under supply constraints, while acquisition strategies must align with supply-demand trajectories, as approaches effective during oversupply periods fail when scarcity is immediate. Release protocols contribute marginally to overall resilience, indicating policy focus should prioritise upstream sourcing and acquisition.