Brightcon 2026, hackathon and courses

Highlights / Discussion Points:

• LLMs can support automated unit process mapping pipelines
• Open-source language models can be fine-tuned to case specific knowledge bases
• These models can be efficiently implemented to open-source software such as Brightway2
• Language model driven methodologies require rigorous validation methods and metrics

Concise Description:
Mitigating environmental impacts requires greater value-chain transparency for companies through product-level disclosure of material origins, transportation, composition, and end-of-life disposal methods. Life cycle assessment (LCA), a systematic approach which quantifies the cradle-to-grave environmental impacts of products, can support improved transparency for product designers and policymakers in this context. Despite providing useful decision-support for manufacturers, LCA faces challenges when deployed at companies with extensive product portfolios. These challenges can be traced back to life cycle inventory (LCI) analysis. Establishing an LCI is the step in which product system boundaries are defined and inputs and outputs to the environment are quantified. This step often involves the selection of a unit process from a database which fits component descriptors usually found in the manufacturer’s bill of materials (BoM) and the decision of an LCA practitioner. Our algorithm was engineered to assist practitioners with this time-consuming task.

Given the interpretative nature of the selection process, Natural language processing (NLP) was identified as an appropriate decision-making tool. A subfield of computer science and artificial intelligence, NLP uses machine learning to enable computers to “understand” and communicate human language. Our algorithm enables users to input product BoMs, then returns a full cradle-to-gate midpoint impact report. We achieve this workflow by using a large language model’s (LLM) (GPT 4o-mini) capability to translate domain-specific descriptors to natural language and open-source Hugging Face models for unit process retrieval. All the unit process metadata extraction, technosphere and biosphere matrix construction and midpoint impact calculations are handled using Ecoinvent 3.12 .xml files and Brightway2 python functions. All codes will be made publicly available with open-source LLM Ollama variants, making the presented methods reproducible locally.

This collaborative research between academia and industry, including multiple divisions of Schneider Electric, supports data transparency at scales relevant for multinational firms with broad-ranging, diversified product portfolios.