A comparative study on machine learning models for estimating cathode material compositions in black powders using prompt gamma-ray spectra

According to the rapid growth of the Lithium-ion battery (LIB) recycling industry, the demands for rapid and non-destructive analysis of black powder, the key product in this recycling process, has increased to enable effective recovery of valuable materials and the formulation of optimal recycling strategies. This study aims to evaluate the feasibility of a deep learning application that estimates the composition of valuable materials in black powder in prompt gamma-ray activation analysis (PGAA). Reference data for major target materials, Nickel, Cobalt, Manganese, and Iron, were acquired using the Four-circle single-crystal diffractometer at the HANARO research reactor, and synthetic datasets with varying counting statistics were generated for model training. To identify the most suitable model for PGAA-based composition estimation, we conducted a comparative analysis of six machine learning algorithms, including machine learning (Lasso, Decision Tree, XGBoost) and deep learning architectures (MLP, CNN, Transformer). Among them, the Transformer architecture demonstrated best performance, achieving an r2 score of over 0.99 for all components. Furthermore, uncertainty quantification using deep ensembles confirmed that the model provides reliable confidence intervals, essential for industrial decision-making. These results demonstrate the promising potential of the deep learning method for effective and high-throughput battery recycling applications.