Machine Learning Approaches for Thermoelectric Performance Predictions

Abstract

The area of thermoelectric (TE) research suffers from an affordable pathway to achieve high performance TE materials. This is because the merit of the experimental approach, although sacrosanct and irrefutable, often, resorts to a trial- and-error method approach. This approach is achieved through training, experience, and observed knowledge. Additionally, while working to find an effective solution through this approach, the target TE material is kept in mind. This introduces biasness in TE materials discovery. In TE research, recent studies have demonstrated the potential for accelerated materials discovery through artificial intelligence (AI) driven methods. Building on these advances, the thesis aims to assist experimental researchers in predicting the properties for high-performance thermoelectric (TE) materials. The objective in the thesis is realized with the developed and tested machine learning (ML) models to predict TE properties. The developed models are based on extreme gradient boosting (XGBoost) and light gradient boosting machine (LightGBM) algorithms. These algorithms, which form part of the ML framework, were trained using curated datasets. The models achieved good predictive accuracy of TE properties. The interpretability of the model predictions through SHapley Additive exPlanations (SHAP), provided interesting and chemically meaningful insights between TE material compositions and the TE properties. In one of our studies, the predictive models were validated experimentally for new doped SnSe systems to observe near consistency between predicted and measured κ values. Subsequently, we developed an end-to-end web application embedded with these ML models, hosted on Git-Hub and Microsoft Azure cloud to deliver rapid TE property predictions. The application is made accessible to TE researchers worldwide. The researchers can upload any set of compositions to the web interface and receive immediate thermoelectric (TE) property predictions. The methodology of the overall ML pipeline explained in the chapters are open for diversification with other Deep Learning (DL) algorithms or Generative AI (Gen AI) models. Furthermore, the ML models can be retrained by modifying the data in the existing dataset, in the direction of improving model accuracy. With this, the models can be used for experimental or first-principle based computational validation. The scope of this research offers more questions than answers leading to an extensive scope of hypothesis generation. Therefore, this opens unlimited opportunities for future investigations.