[Paper Review] Unlocking Thermoelectric Potential: A Machine Learning Stacking Approach for Half Heusler Alloys

Vipin K. E, Prahallad Padhan|arXiv (Cornell University)|Aug 1, 2024

Heusler alloys: electronic and magnetic propertiesMaterials Science3 citations

TL;DR

This study proposes a machine learning stacking model combining Random Forest and XGBoost to predict thermoelectric properties of Half Heusler alloys with high accuracy. The ensemble approach achieves superior performance over individual models, identifying temperature, mean covalent radius, and average Gibbs energy deviation as key predictors of ZT, enabling data-driven design of high-performance thermoelectric materials.

ABSTRACT

Thermoelectric properties of Half Heusler alloys are predicted by adopting an ensemble modelling approach, specifically the stacking model integrated using Random Forest and XGBoost scheme. Leveraging a diverse dataset encompassing thermal conductivity, the Seebeck coefficient, electrical conductivity, and the figure of merit (ZT), the study demonstrates superior predictive performance of the stacking Model, outperforming individual base models with high R2 values. Key features such as temperature, mean Covalent Radius, and average deviation of the Gibbs energy per atom emerge as critical influencers, highlighting their pivotal roles in optimizing thermoelectric behavior. The unification of Random Forest and XGBoost in the stacking model effectively captures nuanced relationships, offering a holistic understanding of thermoelectric performance in Half Heusler alloys. This work advances predictive modelling in thermoelectricity and provides valuable insights for strategic material design, paving the way for enhanced efficiency and performance in thermoelectric applications. The ensemble modelling framework, coupled with insightful feature selection and meticulous engineering, establishes a robust foundation for future research in pursuing high-performance thermoelectric materials.

Motivation & Objective

• To develop a robust predictive model for thermoelectric properties in Half Heusler alloys using machine learning.
• To improve prediction accuracy beyond individual models by integrating multiple estimators through ensemble stacking.
• To identify the most influential material descriptors governing thermoelectric performance, particularly ZT.
• To enable data-driven design of high-efficiency thermoelectric materials through feature importance analysis.
• To establish a scalable framework for predictive materials discovery in complex intermetallic systems.

Proposed method

• An ensemble stacking model is constructed by combining predictions from base estimators: Random Forest and XGBoost.
• The meta-learner uses the predictions of base models as input features to generate final output predictions for ZT, Seebeck coefficient, electrical conductivity, and thermal conductivity.
• A comprehensive dataset including temperature, mean covalent radius, and Gibbs energy per atom is used for training and feature engineering.
• The model is trained and validated using regression metrics, with R² used to evaluate predictive performance.
• Feature importance analysis is performed to identify critical descriptors influencing thermoelectric behavior.
• The framework integrates hyperparameter tuning and cross-validation to ensure robustness and generalization.

Experimental results

Research questions

• RQ1 Can a stacking ensemble model outperform individual machine learning models in predicting thermoelectric properties of Half Heusler alloys?
• RQ2 Which material descriptors—such as temperature, covalent radius, or Gibbs energy deviation—most significantly influence ZT?
• RQ3 How do the interactions between structural and thermodynamic features affect thermoelectric performance?
• RQ4 To what extent does the integration of Random Forest and XGBoost enhance predictive accuracy compared to standalone models?
• RQ5 Can the model provide actionable insights for the rational design of high-ZT Half Heusler materials?

Key findings

• The stacking model achieved higher R² values than individual base models, demonstrating superior predictive performance for ZT and other thermoelectric properties.
• Temperature, mean covalent radius, and average deviation of Gibbs energy per atom were identified as the most influential features in determining ZT.
• The integration of Random Forest and XGBoost effectively captured complex, non-linear relationships in the data, improving generalization and robustness.
• The model's high predictive accuracy enables reliable screening of Half Heusler alloys for thermoelectric applications.
• Feature importance analysis revealed that thermodynamic stability (via Gibbs energy deviation) and atomic size (via covalent radius) are critical for optimizing ZT.
• The framework provides a scalable, interpretable, and high-performance approach for accelerating the discovery of next-generation thermoelectric materials.