Toward efficient dielectric material discovery: a comparative evaluation of Transformer and AutoML-based inverse learning

The discovery of dielectric material compositions and processing parameters often relies on trial-and-error experimentation, which is time-consuming, costly, and labor-intensive. To accelerate this process, artificial intelligence (AI) provides data-driven solutions that enhance predictive accuracy while minimizing experimental overhead. This study compares two AI methodologies: Transformer and AutoML-based Inverse Learning, for recommending compositions and processes of dielectric materials using data from the https://www.matcenter.org/Ceramic Materials Information Bank. Results show that while both methods perform reliably, AutoML consistently outperforms Transformer models across all metrics. AutoML achieved RÂ2 scores of 0.970 for component prediction and 0.978 for process prediction, with MAPE values of 1.4% and 1.1%, respectively. In contrast, Transformer models yielded lower RÂ2 scores (0.952 and 0.934) and higher MAPE values (over 2.0%). To enhance model interpretability, SHAP (SHapley Additive exPlanations) analysis was conducted, providing insights into feature contributions and supporting the reliability of predictions. These findings demonstrate AutoML superior ability to handle tabular data and its effectiveness in capturing complex material-process relationships. The study establishes AutoML as a robust framework for accelerating material discovery. It highlights the importance of selecting architectures that are suited to the data structure in AI-driven materials science.