TL;DR
This paper introduces a novel multi-agent system that quantifies the carbon footprint of global supply chains, including the often-overlooked AI compute costs. It develops a "carbon-adjusted supply chain performance" (CASP) metric to assess trade-offs between service levels and emissions, offering actionable insights for optimizing logistics under carbon constraints.
What Problem Does This Paper Address?
Global supply chains are increasingly complex and face dual pressures: maintaining operational resilience amidst disruptions and meeting stringent decarbonization goals. While AI is being adopted to optimize these chains, its own energy and carbon footprint (the "carbon cost of intelligence" or CCI) has been largely ignored. Furthermore, existing carbon assessments often aggregate data, masking critical differences between product categories with varying constraints (e.g., cold-chain requirements for pharmaceuticals vs. standard logistics for clothing). This paper addresses the need for a holistic, package-type-specific framework that integrates physical logistics, AI compute, and national energy grid variations to enable truly sustainable and resilient supply chain design.
Methods
The core of this work is a multi-agent system designed to model supply chain dynamics. This system features:
- LLM-powered Agents: Orchestrator, Risk Agent, and Sourcing Agent, leveraging large language models (LLMs) for reasoning and decision-making.
- Tool-Augmented Reasoning: Agents interact with deterministic analytical engines (e.g., predictive models, carbon calculators) via structured tool interfaces, enabling them to perform quantitative analysis rather than just text retrieval.
- Carbon-Adjusted Supply Chain Performance (CASP) Metric: A novel metric that integrates physical transport carbon, cold-chain overhead, and AI compute carbon into a single, per-package-type performance measure.
- Predictive Analytics: Gradient Boosting models are used for predicting delay probability, on-time performance, and cost.
- Vendor Segmentation: K-Means clustering is employed to categorize carriers based on cost and on-time performance.
- Carbon Cost of Intelligence (CCI) Layer: Quantifies the energy and carbon emissions associated with AI model inference.
- Early-Warning Indicators: A system for identifying potential disruption amplification risks (supplier concentration, geographic clustering, cold-chain fragility).
Datasets
- Indian Logistics Dataset: A 25,000-record dataset containing features like package type, origin, destination, distance, vehicle type, delivery partner, and delay rate. This dataset was used for training ML models and evaluating the system.
- Reference Data: Includes carrier profiles, city and route information, grid carbon intensity data (from EPA eGRID, Electricity Maps), and AI model energy consumption data.
Key Insights and Ideas
- Package-Type Heterogeneity is Crucial: The "stakes tier" (critical, high-value, standard) and specific constraints like cold-chain requirements dramatically impact optimization potential and carbon footprint. Critical logistics offer limited optimization (<5%), while standard logistics can achieve significant reductions (up to 40%).
- AI Compute Carbon is Small but Growing: While AI compute carbon is orders of magnitude smaller than transport carbon for a single shipment, it becomes relevant at scale or when considering placement in high-carbon grids. This highlights the importance of carbon-aware AI inference.
- Governance Levers Have Asymmetric Impacts: Sourcing rules (e.g., preferring regional suppliers) and buffer policies offer the most significant carbon reductions, especially for standard package types. Compute policies have a minimal impact on total carbon for single shipments.
- Socio-Technical System View is Essential: Supply chain resilience and sustainability are emergent properties of interactions between physical assets, digital infrastructure, and national energy systems. Optimizing one component in isolation can shift burdens to others.
Key Results
- Delay Prediction: Achieved an F1 score of 0.954 ± 0.003 and accuracy of 0.976 for delay classification.
- On-Time Prediction: Achieved accuracy of 0.976 ± 0.002 and F1 of 0.984 ± 0.002.
- CASP Metric: Demonstrated significant variation in CASP across package types, from 6.67 × 10⁻⁴ for pharmacy (critical, cold chain) to 17.13 × 10⁻⁴ for furniture (standard).
- Carbon Optimization: Critical types showed ~4% optimization potential, while standard types showed ~40%.
- AI Compute Carbon: Showed a 20x variation between Norway (low grid intensity) and India (high grid intensity) for AI compute per inference.
- Pipeline Runtime & Cost: LLM agent calls dominate pipeline runtime (15-30s per agent) and cost (~$0.05-$0.10 per query).
Limitations and Future Work
- Dataset Specificity: Findings are grounded in a single Indian logistics dataset and carrier ecosystem; generalization to other geographies requires substituting region-specific inputs.
- Transport Carbon Model: Relies on distance-based per-vehicle emissions, with implicit assumptions about loading and no explicit representation of backhaul.
- AI Compute Model: Uses literature-based median energy-per-inference values and a small number of LLM calls; high-compute deployments or different hardware could yield higher CAI.
- Future Work:
- Incorporate Scope 3 emissions from supplier operations.
- Implement real-time carbon optimization using marginal grid intensity forecasts.
- Develop portfolio-level aggregation methods.
- Validate across additional datasets and geographies.
- Extend CASP to include other emissions and optimize for real-time grid intensity.
Summary updated