Robust Machine Learning Framework for Reliable Discovery of High-Performance Half-Heusler Thermoelectrics

TL;DR

This work addresses the challenge of poor generalizability in ML-driven discovery of thermoelectric materials by proposing a robust workflow tailored to half-Heusler systems. It combines a PCA-based train/test split to meaningfully sample chemical space, Bayesian hyperparameter optimization with $k$-best feature filtering across multiple models, and the SISSO framework for physically interpretable descriptors, complemented by SHAP analyses. A high-throughput screening pipeline with database-informed stability filters, temperature interpolation to a target condition, and ensemble averaging yields several novel high-$zT$ candidates, illustrating the practical impact of improved generalizability over mere metric optimization. The approach provides a scalable, interpretable blueprint for data-driven design of next-generation thermoelectrics, with potential integration to thin-film synthesis and active-learning validation strategies.

Abstract

Machine learning (ML) can facilitate efficient thermoelectric (TE) material discovery essential to address the environmental crisis. However, ML models often suffer from poor experimental generalizability despite high metrics. This study presents a robust workflow, applied to the half-Heusler (hH) structural prototype, for figure of merit (zT) prediction, to improve the generalizability of ML models. To resolve challenges in dataset handling and feature filtering, we first introduce a rigorous PCA-based splitting method that ensures training and test sets are unbiased and representative of the full chemical space. We then integrate Bayesian hyperparameter optimization with k-best feature filtering across three architectures-Random Forest, XGBoost, and Neural Networks - while employing SISSO symbolic regression for physical insight and comparison. Using SHAP and SISSO analysis, we identify A-site dopant concentration (xA'), and A-site Heat of Vaporization (HVA) as the primary drivers of zT besides Temperature (T). Finally, a high-throughput screening of approximately 6.6x10^8 potential compositions, filtered by stability constraints, yielded several novel high-zT candidates. Breaking from the traditional focus of improving test RMSE/R^2 values of the models, this work shifts the attention on establishing the test set a true proxy for model generalizability and strengthening the often neglected modules of the existing ML workflows for the data-driven design of next-generation thermoelectric materials.

Figures (8)

• Figure 1: a) Illustration in 2D of the ordering of the materials by increasing distance from the centroid of the PCA space. b) Distribution of the materials in the training (Tr) or test (Te) sets for all the folds, based on the order defined in a).
• Figure 2: Two-dimensional t-SNE projection of each fold's test set
• Figure 3: Ranking of the elemental features by absolute Pearson coefficient
• Figure 4: (a) Pearson correlation matrix of five sample elemental features of the A-site elements; (b) Evolution of the number of features $k_\mathrm{max}$ retained as a function of the $r_{\mathrm{min}}$.
• Figure 5: a) Optimization of the XGBoost hyper-parameters with the features set by the pair $(r_{\mathrm{min}}, k) = (0.4, 41)$; b) $k$-best selection: the pair $(r_{\mathrm{min}}, k) = (0.4, 41)$ leads to a validation RMSE of $0.152$; c) Effect of $k_\mathrm{best}$ over the use of $k_\mathrm{max}$: decrease of the validation RMSE and of the threshold $r_{\mathrm{min}}$ leading to a decreased number of features.