Predictive modeling of metal recovery efficiencies from spent lithium-ion batteries using RSM enhanced residual neural network framework

The increasing consumption of lithium-ion batteries (LIBs) has created a growing need for efficient recycling processes to recover valuable metals such as lithium (Li) and cobalt (Co). The present study analyses the performance of supervised machine learning models in predicting the leaching efficiencies of Li and Co from spent LIBs in the ultrasound-assisted leaching of Lithium Cobalt Oxide (LCO) cathode material using a Hybrid MW-US high-pressure chemical reactor. The standalone linear regression (LR), quadratic polynomial regression (QPR), neural network (NN) and random forest (RF) models were firstly evaluated by their ability to capture the relationship between key process parameters and the metal recovery efficiencies. Among the standalone models, the RF model demonstrated the strongest predictive performance. However, these purely data-driven models do not explicitly incorporate the empirical process relationships identified from the experimental leaching process. To address this limitation, an RSM-enhanced residual neural network (RSM-NN) framework was developed by integrating the response surface model (RSM) formulation from the original experimental study as an empirical process constraint during training. The proposed RSM-NN framework achieved very high predictive accuracy and produced smooth and consistent trends in the partial dependency analysis, reflecting the strength of the hybrid modeling framework. The results demonstrate the potential of RSM-enhanced machine learning frameworks for improved modelling and optimization of hydrometallurgical recovery processes from spent LIBs under the investigated experimental conditions.