FEED: Constraint-Aware Machine Learning for Climate-Impact-Aware Surplus Food Diversion and Methane Mitigation

Food loss and waste generate 8–10% of global greenhouse gas emissions, nearly five times the emissions of the entire aviation sector, while 783 million people face hunger. Satellite monitoring has repeatedly identified individual landfills, including Delhi's Ghazipur site, among the world's largest methane “super-emitters.” Yet existing surplus food platforms stop at binary donation-matching and default to landfill the moment a match fails, and computer vision work on food condition stops at state estimation without ever connecting a prediction to a downstream action. We close that gap. We propose FEED (Food-waste Evaluation via Explainable Destination-routing), a two-stage system that (i) formulates destination selection as constrained multi-label classification over donation, animal feed, compost, and anaerobic digestion, gated by a concept-based safety mechanism evaluated at projected arrival time rather than pickup time, and (ii) formulates the downstream facility assignment as a learned routing problem. We ground the destination hierarchy in the EPA's 2023 Wasted Food Scale, giving FEED a direct line into existing climate infrastructure: anaerobic digestion into bio-CNG and compost into fertiliser, so that every routed item has a concrete, auditable emissions outcome. This is an early-stage proposal; we lay out FEED's architecture, its safety guarantees, and the open questions on which we are seeking mentorship.