Unveiling Scope 3 emissions in energy supply chains: a graph neural network approach for missing data imputation and optimization

Introduction Accurate accounting of Scope 3 emissions is essential for decarbonization in energy supply chains, but remains difficult because of data opacity and the complexity of multi-tier supplier networks. Methods We formulate Scope 3 propagation using a Leontief-inspired model on a directed supply graph and propose a physics-informed Leontief-GNN with an input-injection mechanism to impute missing firm-level emissions. We further integrate the imputed emission map into a multi-objective supplier selection framework. Results On a synthetic scale-free network of 2,000 nodes, the proposed Leontief-GNN achieved an RMSE of 144.41, outperforming feature-only baselines and a matched GraphSAGE ablation without input injection (RMSE 172.31). The framework also generated Pareto-efficient procurement strategies and showed that high-emitting firms tend to be structural hubs whose targeted failure rapidly degrades network connectivity. Discussion The proposed framework provides a physically grounded and practically useful approach for missing Scope 3 emission imputation and procurement optimization, supporting data-driven decarbonization and ESG-aware supply-chain decision-making in the energy sector.