Multi-Objective Optimization Employing Knowledge Graph-Embedded Large Language Model to Strategize Battery Recycling Technology Selection
Description
This study presents two AI-driven frameworks to support sustainable battery recycling: (1) the Battery-LLaMA module, which integrates a fine-tuned LLaMA-2-7B model with a battery recycling-informed knowledge graph (BR-KG) for domain-specific question answering, and (2) a multi-objective optimization (MOO) framework for process selection based on internal energy consumption at the molecular level and revenue generation from the recycling process. The BR-KG, built from over 10k+ Elsevier abstracts using named entity recognition, captures structured recycling knowledge and enhances the performance of the Battery-LLaMA model (F1-score: 0.821 vs. 0.701 for ChatGPT-4o). The MOO framework reveals that hydrometallurgical processes, particularly those combining leaching, roasting, and regeneration, offer the most economically and energetically favorable pathways for recycling 100 kg of Li-ion batteries with >90% efficiency.
Files
Steps to reproduce
To regenerate this work from scratch, follow the steps below: Step 1: Extract the text (e.g., abstracts of articles) from any reliable source, such as the Elsevier database. Step 2: Annotate the extracted data sample as per the requirement. In this work, we considered two tasks: (i) an entity extraction task, where the following entities were annotated: battery type, recycling process name, process conditions, and recovered rare earth materials (i.e., Li, Ni, Co, Pb, Mn, etc.) in the given text; and (ii) a question-answering task, where a curated set of question and answer pairs was created. These types of annotated datasets were utilized to fine-tune open-source large language models. Step 3: The resulting fine-tuned LLM versions were used to generate the Battery Recycling-Informed Knowledge Graph (BR-KG), which was then integrated with a fine-tuned question-answering model named Battery-LLaMA. Step 4: In the final step, we formulated a multi-objective optimization problem with the goal of minimizing internal energy consumption and maximizing revenue from the recycling process. This resulted in a Pareto plot that illustrates the trade-offs between the two objectives.
Institutions
- Arizona State University
- Indian Institute of Technology Delhi