Towards Autonomous Sustainability Assessment via Multimodal AI Agents

Zhihan Zhang, Alexander Metzger, Yuxuan Mei, Felix Hähnlein, Zachary Englhardt, Tingyu Cheng, Gregory D. Abowd, Shwetak N. Patel, Adriana Schulz, Vikram IyerarXiv 2025 · DOI · arXiv

Introduction

The environmental impact (EI) of consumer electronics is a critical concern for both regulators and consumers. However, the standard methodology for quantifying this impact—Life Cycle Assessment (LCA)—is plagued by significant bottlenecks. Traditional LCAs require a detailed Life Cycle Inventory (LCI), which involves manually cataloging every component, material, and manufacturing process. This process is notoriously slow, often taking weeks or months, and relies heavily on proprietary supply chain data that is frequently inaccessible. Furthermore, human experts must often rely on inconsistent proxies when data is missing, leading to high variability in results.

Methodology: The Auto Multi-AI Agent LCA

To address these challenges, this research introduces an autonomous multi-agent system designed to perform LCAs using only publicly available data. The system reimagines LCA as a hierarchical information retrieval and reasoning task.

Overview of traditional vs. autonomous multi-agent LCA frameworks, including information retrieval and impact assessment modules.

Autonomous Multi-Agent Self-Play

The core of the system is a "self-play" interaction between two AI agents:

  • LCA Agent: Acts as the architect and critic, defining the assessment scope and critiquing the generated inventory.
  • Stakeholders Agent: Functions as a data retriever, utilizing various tools to find missing information from online and proprietary sources.

Multimodal Information Retrieval

The system employs a sophisticated multimodal toolchain to extract data from visual and textual sources:

  • Visual Mining: Using YOLO-based object detection, the system identifies components on printed circuit boards (PCBs). It utilizes a Fast Fourier Transform (FFT) filter to automatically select the highest-quality technical images from search results, outperforming standard Vision-Language Models (VLMs) in technical image selection.
  • Dimension Calibration: Physical sizes are estimated by cross-referencing "anchor" components with known dimensions to calibrate pixel-to-millimeter ratios.

Impact Estimation

For rapid assessment, a kNN-Weighted Gaussian Estimator maps product specifications to clusters of similar products with known footprints. This provides an estimate with built-in uncertainty bounds. Additionally, the system uses material properties (e.g., melting points) to mathematically derive emission factors for unknown materials, reducing reliance on human-selected proxies.

Key Results

The proposed system demonstrates significant improvements over traditional manual practices across multiple performance axes.

System performance and user perception comparison between traditional LCA and the proposed AI-driven approach.
  • Efficiency: The system reduces the time required for an LCA from months to under 60 seconds.
  • Accuracy: Compared to official carbon footprint reports from major manufacturers (Apple, Dell, etc.), the end-to-end system achieved a Mean Absolute Percentage Error (MAPE) of 18.02%. The lightweight kNN model achieved an even lower MAPE of 12.28% for laptops and desktops.
  • Expert Benchmark: In a direct comparison, the AI agent outperformed nine human experts in selecting correct emission factors for masked materials, reducing error rates by over 120%.

Discussion and Conclusion

A primary insight of this work is that proprietary data is not a prerequisite for high-fidelity sustainability assessments. By leveraging the "visual and textual exhaust" found in government certifications (e.g., FCC reports) and enthusiast communities (e.g., iFixit), AI agents can reconstruct accurate inventories.

Furthermore, the system's design prioritizes explainability; the kNN-Gaussian approach allows auditors to trace which reference products influenced a specific estimate. While limitations remain—such as challenges with irregular PCB geometries and the need for alignment with specific corporate accounting boundaries—this research paves the way for scalable, transparent, and autonomous sustainability science. Future work will explore "Agentic Review" processes to further enhance the reliability of these autonomous assessments.

Summary updated

BibTeX
@misc{zhang2025towards,
  author = {Zhihan Zhang and Alexander Metzger and Yuxuan Mei and Felix Hähnlein and Zachary Englhardt and Tingyu Cheng and Gregory D. Abowd and Shwetak N. Patel and Adriana Schulz and Vikram Iyer},
  title = {{Towards Autonomous Sustainability Assessment via Multimodal AI Agents}},
  year = {2025},
  eprint = {2507.17012},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2507.17012},
}

Loading related content...

Figure