Explainable AI for Material Property Prediction Based on Energy Cloud: A Shapley-Driven Approach

The scientific community has raised increasing apprehensions over the transparency and interpretability of machine learning models employed in various domains, particularly in the field of materials science. The intrinsic intricacy of these models frequently results in their characterization as "black boxes", which poses a difficulty in emphasizing the significance of producing lucid and readily understandable model outputs. In addition, the assessment of model performance requires careful deliberation of several essential factors. The objective of this study is to utilize a deep learning framework called TabNet to predict lead zirconate titanate (PZT) ceramics' dielectric constant property by employing their components and processes. By recognizing the crucial importance of predicting PZT properties, this research seeks to enhance the comprehension of the results generated by the model and gain insights into the association between the model and predictor variables using various input parameters. To achieve this, we undertake a thorough analysis with Shapley additive explanations (SHAP). In order to enhance the reliability of the prediction model, a variety of cross-validation procedures are utilized. The study demonstrates that the TabNet model significantly outperforms traditional machine learning models in predicting ceramic characteristics of PZT components, achieving a mean squared error (MSE) of 0.047 and a mean absolute error (MAE) of 0.042. Key contributing factors, such as d33, tangent loss, and chemical formula, are identified using SHAP plots, highlighting their importance in predictive analysis. Interestingly, process time is less effective in predicting the dielectric constant. This research holds considerable potential for advancing materials discovery and predictive systems in PZT ceramics, offering deep insights into the roles of various parameters.