⚡ electrical engineering

Explainable Multi-Fidelity Regression for Lithium-Ion Battery Recycling: Maximizing Economic and Environmental Outcomes under U.S. Standards

This paper introduces an explainable multi-fidelity regression framework that integrates PCA-based feature selection with Bayesian Ridge Regression to accurately predict lithium-ion battery recycling circularity, costs, and profitability under U.S. policy scenarios, offering a scalable decision-support tool for optimizing economic and environmental outcomes.

Original authors: Shahriar Siddique Ayon, Md AKbar Hossain, Md. Jisun Abedin Aurnob, Majharul Islam Shanto, Md. Ebrahim Hossain

Published 2026-10-01

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Shahriar Siddique Ayon, Md AKbar Hossain, Md. Jisun Abedin Aurnob, Majharul Islam Shanto, Md. Ebrahim Hossain

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Every day, millions of electric vehicles and portable electronics rely on lithium-ion batteries to store energy. As these devices reach the end of their useful lives, they become a growing mountain of waste, posing serious risks to soil and water if simply discarded. The solution lies in recycling, a process that not only prevents environmental harm but also recovers valuable materials like lithium, cobalt, and nickel, turning waste back into resources for new batteries. However, deciding how to recycle these batteries is a complex puzzle. Different methods exist, each with its own costs, efficiency levels, and environmental impacts. Policymakers and industry leaders need to know which approach will be most profitable and sustainable as regulations change and technology evolves, but predicting these outcomes has historically been difficult without clear, data-driven tools.

A team of researchers has developed a new way to solve this puzzle by creating a smart computer system that predicts the future of battery recycling with high precision. Instead of relying on static estimates or theoretical guesses, this system uses a method called multi-fidelity regression. Imagine trying to map a vast landscape: you might first use a quick, low-resolution sketch to get the general shape of the terrain, and then zoom in with a high-resolution camera to capture the fine details of specific areas. The researchers applied this same logic to their computer model. They combined a fast, broad-stroke predictor with a slower, highly detailed one. This hybrid approach allows the system to process vast amounts of information quickly while still delivering the accuracy needed for critical decisions. The model was trained on three distinct sets of data covering the years 2020 to 2050, focusing on how well materials can be recovered, how much it costs to process them, and how much profit can be made under various U.S. policy scenarios.

The results of this study are striking. When tested against other common prediction methods, the new system proved far superior, achieving a level of accuracy that explains 99% of the variation in the data. For predicting how circular the recycling process is—meaning how much material is successfully recovered and reused—the model was exceptionally accurate. It also provided highly accurate forecasts for processing costs and potential profits, minimizing the mean absolute error to 0.01 for circularity, 0.04 for cost, and 0.11 for profit. To ensure that these predictions were not just black-box guesses, the researchers used a technique called explainable artificial intelligence. This allowed them to pull back the curtain and see exactly which factors were driving the results. They found that the amount of material being recycled, the value of the raw materials being recovered, and the specific type of recycling method used were the primary drivers of success.

The analysis revealed clear patterns that could guide future decisions. For instance, the study showed that a specific policy scenario focused on sustainable development consistently led to better material recovery and higher efficiency compared to other scenarios. When looking at costs, one method known as hydrometallurgy, which uses chemical solutions to extract materials, emerged as the most cost-effective option. However, another method called direct recycling, which preserves the battery's internal structure, was found to be the most profitable in the long run, despite higher initial processing costs. The researchers also discovered that transportation plays a crucial role; moving batteries within the United States by train or truck was less profitable than shipping them to China, though the latter came with a higher carbon footprint. By balancing these economic and environmental factors, the system identified that direct recycling offered the best overall trade-off, maximizing profit while keeping environmental impact relatively low.

To make these insights useful for real-world decision-making, the team built a web-based tool that anyone can use. This interface allows users to input different scenarios, such as changing the type of battery, the location of the recycling plant, or the year of analysis, and instantly see the predicted outcomes for circularity, cost, and profit. This tool acts as a decision-support system for policymakers and industry stakeholders, helping them navigate the complex landscape of battery recycling regulations and market conditions. While the model is currently based on projections and existing data, it offers a scalable and reliable way to optimize recycling strategies. The researchers acknowledge that the system is a proof of concept and would need further testing and security measures before being deployed on a massive industrial scale. Nevertheless, this work provides a powerful new lens through which to view the future of battery recycling, turning a chaotic challenge into a manageable, data-driven opportunity for a more sustainable economy.

Technical Summary: Explainable Multi-Fidelity Regression for Lithium-Ion Battery Recycling

Problem Statement

The rapid proliferation of electric vehicles (EVs) and portable electronics has accelerated the generation of end-of-life lithium-ion batteries (LIBs), creating urgent environmental and economic challenges. While recycling is critical for material recovery and transitioning to a circular economy, there is a significant lack of data-driven tools capable of jointly predicting material circularity, processing costs, and profitability under evolving policy frameworks, specifically the U.S. Recycled Content Standards (RCS).

Existing literature predominantly relies on static material flow analyses or theoretical projections that fail to capture real-world operational heterogeneity and economic trade-offs. Furthermore, current machine learning (ML) approaches often depend heavily on historical data, struggle to predict costs or profits in real-time, and lack interpretability. Consequently, policymakers and industry stakeholders lack a reliable, rapid, and transparent method to forecast the viability of LIB recycling strategies across different future scenarios (2020–2050).

Methodology

The study proposes an Explainable Multi-Fidelity Regression Framework designed to optimize predictions for three key outcomes: material circularity, processing cost, and post-recycling profitability. The methodology follows a structured pipeline:

1. Data Collection and Preprocessing

The research utilizes three curated datasets derived from a publicly available study (Dunn et al., hosted on Dryad), spanning 30 years (2020–2050) with over 800 scenario variations. The datasets are categorized as:

Circularity Dataset: Tracks material flow, energy retirement, and system circularity under various policy/tech scenarios.
Cost Dataset: Breaks down recycling costs ($/kg) by process type, disassembly levels, and transport.
Profit Dataset: Analyzes profitability shifts based on economic and transportation assumptions.

Preprocessing involved handling categorical variables via one-hot and label encoding, outlier detection using the Interquartile Range (IQR) method, and data partitioning (80% training, 20% testing, with 20% of training reserved for validation).

2. Feature Engineering

To identify the most impactful variables, the authors employed a Principal Component Analysis–Information Gain (PCA-IG) hybrid approach. PCA reduced dimensionality, while Information Gain evaluated the relevance of transformed features to the target variables. This process identified key drivers such as recycled weight, demand, recycling process types (hydro, pyro, direct), and virgin material values.

3. Model Architecture: Fidelity-Based Regression

The core contribution is a novel Fidelity-Based Bayesian Ridge Regression (BRR) model. This architecture integrates:

Low-Fidelity (LF) Model: A computationally efficient predictor for rapid initial approximation.
High-Fidelity (HF) Model: A probabilistic BRR component that provides high precision but requires more resources.
Integration: The final prediction ( $\hat{y}$ ) is a weighted combination of both models: $\hat{y} = \alpha f_{LF} + (1-\alpha)f_{HF} + \epsilon$ .
Optimization: The model was optimized using GridSearchCV for hyperparameter tuning (specifically $\alpha$ and $\lambda$ values) and compared against Particle Swarm Optimization (PSO)-tuned models.

4. Explainability and Deployment

To ensure transparency, Local Interpretable Model-agnostic Explanations (LIME) were applied to disaggregate feature influences, identifying positive and negative drivers for each prediction. Finally, a web-based decision-support interface (built with React.js and Flask) was developed to allow stakeholders to input parameters and receive real-time predictions for circularity, cost, and profit.

Key Contributions

PCA-IG Feature Selection: Application of a hybrid feature selection method to filter variables and focus on the most impactful factors influencing circularity, cost, and profit.
Superior Predictive Accuracy: The proposed fidelity-based BRR model outperformed traditional benchmarks (Random Forest, SVR, Linear Regression, KNN, and PSO-optimized models) across all three datasets.
Explainable AI Integration: The use of LIME to identify specific drivers (e.g., recycling throughput, virgin material value) that positively or negatively impact economic and environmental outcomes.
Operational Decision Support: Development of a functional web interface enabling real-time scenario planning for policymakers and industry stakeholders.
Insight into Recycling Methods: Analysis revealing the efficiency trade-offs between hydrometallurgical, pyrometallurgical, and direct recycling methods under U.S. regulatory contexts.

Experimental Results

The proposed model demonstrated superior generalization capabilities across all three datasets:

Circularity Prediction: Achieved an $R^2$ score of 0.99 with a Mean Absolute Error (MAE) of 0.01.
Cost Prediction: Achieved an $R^2$ score of 0.99 with an MAE of 0.04.
Profit Prediction: Achieved an $R^2$ score of 0.99 with an MAE of 0.11.

In benchmarking against other models (including PSO-tuned BRR and SR-BPNN), the Fidelity-Based BRR consistently yielded the lowest error rates and highest $R^2$ scores, particularly after feature selection.

Key Insights from Analysis:

Circularity: The Sustainable Development Scenario (SDS) consistently outperformed the Stated Policies Scenario (STEPS) in material recovery.
Cost: Hydrometallurgical recycling was identified as the most cost-effective method ($8.38/kg), while Direct recycling, though more expensive ($12.41/kg), offered higher material value.
Profit vs. Environment: Direct recycling provided the best balance of profit and environmental impact, whereas Pyrometallurgical routes were found to be the least profitable and most carbon-intensive under U.S. logistics conditions.

Significance and Claims

The paper claims that the proposed framework offers a scalable, interpretable, and deployment-ready solution for optimizing LIB recycling strategies within U.S. regulatory and sustainability contexts. By synergizing computational efficiency with high-precision probabilistic modeling, the framework addresses the gap between theoretical material flow analysis and real-time operational decision-making.

The authors emphasize that the integration of XAI transforms the model from a "black box" into a transparent decision-support system, enabling policymakers to understand the specific levers (e.g., collection rates, transport modes, process types) that drive economic and environmental performance. The study positions this tool as a critical asset for navigating the complex trade-offs required to meet evolving Recycled Content Standards and fostering a circular battery economy.

The authors acknowledge limitations, noting that the model is trained on scenario projections (2020–2050) and specific chemistries (LFP, NMC, NCA), making it primarily a strategic foresight tool rather than a dynamic, real-time operational monitor for emerging battery technologies. Future work is suggested to integrate IoT sensor data and expand to emerging green leaching technologies.