TY  - EJOU
AU  - Jawad, Wasnaa Kadhim 

TI  - Interpretable AI for Non-Destructive Prediction of Electrode Properties from Frequency-Domain Ultrasonic Signals
T2  - Computers, Materials \& Continua

PY  - 
VL  - 
IS  - 
SN  - 1546-2226

AB  - The accuracy of the electrode properties is important in the lithium-ion battery manufacturing process because the thickness variation is a direct influence on the compaction and structural uniformity, transport behavior and overall manufacturing quality. Of the different types of monitoring, ultrasonic frequency-domain relies on a non-destructive pathway for quality evaluation in a process-aware manner and is a promising approach; but, interpretable predictive modeling has been limited at the electrode level. In this study, an open-access database of ultrasonic frequency-domain data of lithium-ion battery electrodes under coating and calendering conditions was used to develop an artificial intelligence (AI) framework for predicting electrode thickness. The data set consisted of two subsets (anode and cathode) each of which included metadata, frequency bins and the associated magnitudes in a spectrum. It involved a combination of preprocessing, feature preparation, Stochastic Gradient Descent optimization of the Deep Neural Network (DNN) training process, and Shapley Additive Explanations (SHAP) for the post hoc analysis of the feature importances. To accommodate differences between the data sets, process variables, and spectral behavior, separate models were created for the anode and cathode sets. The results reported that the proposed framework has a good performance for the anode dataset, where the model got a mean absolute error of 9.530, root mean square error of 16.649, and coefficient of determination of 0.824. The values suggest that the model had a good predictive power on the relationship between ultrasonic frequency-domain input and electrode thickness. The interpretation stage also showed that the presence of certain components of the mid- and high-frequency spectra were important for the prediction of the anode, indicating a relationship between the acoustic response and the variations of material state caused by the manufacturing process. Furthermore, the overall dataset analysis revealed unique distribution in anode and cathode thickness, behavior under calendering state, and spectral-response patterns. The signals of the ultrasonic frequency domain can be effectively combined with deep learning and explainable AI to assist in interpretable thickness prediction, it was concluded. The proposed framework presents an excellent computational tool for non-destructive quality monitoring and intelligent manufacturing control in the production of advanced batteries.
KW  - Battery manufacturing; ultrasonic sensing; thickness prediction; explainable AI; deep neural network

DO  - 10.32604/cmc.2026.084636