Geospatial AI and preference-robust spatial statistics for equity-aware prioritization of U.S. circular food-system infrastructure

National food-waste planning requires identifying where recovery logistics and organics infrastructure warrant closer investigation without implying false precision from fixed rankings. We developed a preference-robust, equity-aware geospatial framework for the United States by integrating 973,904 EPA Excess Food Opportunities Map records with 2020–2024 American Community Survey estimates and 2024 county geography. Six standardized criteria represented excess-food pressure, organic-recipient distance and scarcity, poverty, SNAP participation, and household vehicle access. Stochastic multicriteria rank acceptability propagated EPA low, midpoint, and high generation estimates and 48,000 stakeholder-weight draws across four preference regimes. Among 3219 eligible counties, 322 entered the equal-weight top decile, whereas only 87 retained at least 0.80 top-decile probability across every regime; sensitivity analyses yielded 61–106 core counties. Local Moran analysis identified 254 nominal High–High counties, of which 180 remained High–High after Benjamini–Hochberg adjustment. Although unadjusted prediction residuals were larger in consensus-core counties, distance-conditioned analyses showed that the disparity was driven largely by their concentration at extreme recipient distances rather than an independent core-membership penalty. In a 25-site hypothetical investment screen, the 5% equity Pareto-knee solution reduced national tonnage-weighted organic-recipient distance by 6.33% and consensus-core mean distance by 21.66%, with 24 of 25 selected counties stable across EPA estimate levels. The framework separates robust priorities from preference-sensitive decisions while explicitly representing uncertainty, spatial dependence, equity, and model reliability. Outputs are screening indicators rather than construction-ready recommendations.