Uncertainty in AI-Augmented Climate Risk Analytics for Global Food Supply Chains

About the project

About the Project Background The global food system faces escalating climate risks, where local production shocks (droughts, floods, or heatwaves) cascade through trade networks, threatening food security worldwide. Current models capture network structure but treat human responses as simple heuristics, missing adaptive, cognition-driven behaviours that amplify or dampen cascades. This project merges a validated multiplex shock-propagation framework (Fosch et al., arXiv:2603.01740) with a novel foundation model trained to predict and emulate human decision-making under uncertainty: learning from behavioural data, market reactions, and policy logs. By embedding this AI-driven cognition layer, we will simulate how subjective risk perceptions, strategic delays, and non-rational choices reshape contagion pathways under climate extremes. Impact: actionable uncertainty maps for resilient food supply chains and trade policies, early warning of cascading failures, and AI-augmented stress testing for global supply chains. PhD project While state-of-the-art shock-propagation models accurately map trade-network vulnerabilities to extreme weather events, they universally treat human responses (export bans, hoarding, strategic delays) as static heuristics, ignoring the adaptive, subjective, and often non-rational cognition that fundamentally reshapes contagion pathways under escalating climate extremes. The pivotal question is not merely which nodes fail, but how and when decision-makers perceive risk and act under deep uncertainty, and how these cognitive feedbacks amplify or dampen cascades across global food supply chains. This project tackles this by applying a foundation model, trained on behavioural experiments to emulate human decision-making in real time. You will embed this AI-driven cognition layer into a validated multiplex network framework (Fosch et al., 2026), transforming static topology into a dynamic, behaviour-aware stress-testing engine. Your technical programme spans three pillars: (1) curating and harmonising heterogeneous behavioural, trade, and climate datasets; (2) fine-tuning a transformer-based foundation model to predict subjective risk perceptions and strategic delays under uncertainty; and (3) running large-scale simulations under compound climate extremes to map non-linear cascade effects. Applicant Profile We are seeking highly motivated STEM graduates with excellent programming skills (Python, with familiarity in PyTorch/JAX or network libraries such as NetworkX/igraph). Prior experience in food systems or climate policy is not required, but intellectual curiosity about human-environment interactions and a passion for tackling grand societal challenges are essential.

Research areas