Enhancing Rural Supply Chains with AlphaFold-Inspired Deep Learning
PI: Soundar Kumara (Industrial Engineering)
Plan for funding tuition for graduate students, or the remainder of the researcher’s salary for postdoc and research faculty: 2 Graduate Students through Pearce Chair funding.
Rural supply chains, especially those serving U.S. agriculture and food systems are vital to the national economy. For example, agriculture and related industries contributed roughly $1.537 trillion (5.5% of GDP) in 2023. Including all downstream effects, the broader farm-to-market supply chain has an economic impact on the order of $9.5 trillion (about one-third of GDP) and directly supports ~24 million jobs (~15% of U.S. employment). These chains connect farmers, processors, and retailers across vast distances, underpinning food security and rural livelihoods. Yet rural networks face acute challenges:
1. Infrastructure gaps: Aging roads, bridges, and limited broadband hinder transportation and information flow. For instance, over 70% of U.S. freight (worth $10 trillion) moves by road, but only ~50% of rural roads are rated in good condition.
2. Fragmented coordination: Many small farms and local businesses operate. independently. It is hard to synchronize production, storage, and distribution at scale
3. Volatility and shocks: Agricultural yields are highly variable; extreme weather, pests, and market swings can suddenly disrupt flows. Recent disruptions (COVID lockdowns, crop shortages) illustrated the fragileness of food chains.
4. Data scarcity: Rural areas lack sensors or reliable connectivity for real time analysis.
AlphaFold is a protein structure prediction model developed by DeepMind that achieved a breakthrough, leading to John Jumper and Demis Hassabis winning the Nobel Prize in Chemistry. It primarily uses evolutionary information from other known protein structures to create a larger structure called a multiple sequence alignment (MSA). By supplying the model with an MSA of thousands of homologous proteins, the network starts with a powerful prior: residues that co-evolve are highly likely to make contact in 3D space. The structure prediction module then iteratively refines the representation against this prior, dramatically shrinking the conformational search space and guiding the model toward the residue pairs that matter most. In the rural supply chain setting, we can create an analogous structure: stacks of historical episodes that share similar exogenous drivers (e.g., drought + harvest surplus + limited trucking capacity). Feeding these aligned episodic traces into our attention-based GNN gives the model a learned prior over how the network typically reorganizes under such coupled stresses, allowing it to converge faster, generalize better, and issue earlier warnings.
Building on this analogy, we propose to develop AlphaFold-inspired GNN models for rural supply chain optimization. Modeling the rural supply chain as a dynamic graph, where nodes represent supply chain entities (manufacturers, suppliers) and edges represent logistics links. Nodes could include production capacity or inventory, and edge features include transportation capacity or lead times. Using transformers (attention layers) to “focus” on critical nodes and links when making predictions, just as AlphaFold’s attention focuses on important residue pairs, will allow the network to model long-range effects (e.g. disruption propagation). The model can be trained on historical supply chain data to predict key metrics: future supply/demand balances, risk of delays/shortages, and potential bottlenecks.
This project promises to substantially enhance rural supply chain resilience and efficiency. By applying AlphaFold-like deep learning to supply networks, we aim for early prediction of disruptions (e.g. weather-induced delays or demand spikes) to allow proactive mitigation (e.g. reallocation of inventory, alternate sourcing). An attention-GNN can update predictions as new data arrives, enabling near realtime adaptation to changing conditions.
This project aligns with ICDS mission of promoting cutting edge research. Roar supercomputing infrastructure (expected NVIDIA A100/H100 GPUs, 80GB memory), with distributed training applied as needed. The surrogate will operate on GPU to efficiently handle large 3D tensors and enable fast transitions in the simulated environment, while the lightweights policy network will also reside on GPU to support high-throughput parallel simulations.
2 graduate students for an year will be needed, whom the PI has already identified.