Accelerating Discovery of Vacancy Ordered 18-Valence Electron Half-Heusler Compounds - A Synergistic Approach of Machine Learning and Density Functional Theory

Aug 6, 2026

Upload by Himbo

Vacancy ordered half Heusler compounds with an 18 valence electron count (VHH) are modeled by deriving Ti0.75NiSb, Zr0.75NiSb, and Hf0.75NiSb from 19 VEC precursors. The ordered-vacancy phase is motivated to enable semiconducting behavior while disrupting phonon transport to lower thermal conductivity and increase thermoelectric performance. A machine-learning model trained on 4684 compounds in the Materials Project predicts formation energy (ΔHf), then DFT validates selected candidates, yielding comparable results and dynamical stability via positive phonon modes. Electronic structure and lattice dynamics indicate narrow band-gap semiconductors, and Seebeck, conductivity, thermal conductivity, power factor, and ZT are analyzed.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages5

Reading Time13 min

AccessFree to download

Abstract

Introduction

Hybrid ML+DFT Approach

Accelerating Discovery of Vacancy Ordered 18-Valence Electron Half-Heusler Compounds: A Synergistic Approach of Machine Learning and Density Functional Theory Gowri Sankar S 1,2 , Mukesh K. Choudhary1,2 , Amal Raj V2 and P. Ravindran, a)1,2 1Department of Physics, Central University of Tami Nadu, Thiruvarur, Tamil Nadu, India. 2Simulation Centre for Atomic and Nanoscale MATerials (SCANMAT), Central University of TamilNadu, Thiruvarur, Tamil Nadu 610101, India a) Corresponding author: [raviphy@cutn.ac.in](raviphy@cutn.ac.in) Abstract. In this study we attempted to model vacancy ordered half Heusler compounds with 18 valence electron count (VHH) derived from 19 VEC compounds such as TiNiSb such that the compositions will be Ti0.75NiSb, Zr0.75NiSb and Hf0.75NiSb with semiconducting behavior. The main motivation is that such a vacancy ordered phase not only introduce semi conductivity but also it will disrupt the phonon conducting path in HH alloys and thus reduce the thermal conductivity and as a consequence enhance the thermoelectric figure of merit. In order to predict the formation energy (ΔHf) from composition and crystal structure we have used 4684 compounds for their ΔHf values are available in the material project database and trained a machine learning model with R2 value of 0.943 . Using this trained model, we have predicted the ΔHf of a list ofVHH. From the predicted database of VHH we have selected Zr0.75NiSb and Hf0.75NiSb to validate the machine learning prediction using accurate DFT calculation. The calculated ΔHf for these two compounds from DFT calculation are found to be comparable with our ML prediction. The calculated electronic and lattice dynamics properties show that these materials are narrow band gap semiconductors and are dynamically stable as their all-phonon dispersion curves are having positive frequencies. The calculated Seebeck coefficient, electrical conductivity as well as thermal conductivity, power factor and thermoelectric figure of merit are analyzed. INTRODUCTION Machine Learning (ML) serves various purposes in numerous fields, such as business, medical sciences, and agriculture, among others. The continuous development of advanced algorithms and the availability of vast amount of data have significantly enhanced the accuracy and robustness of ML models. However, further progress is still ongoing to create field-specific or target/property-specific models those can achieve desirable accuracy [1] . Despite the abundance of data, a considerable portion remain unstructured and unclean, making it time-consuming to filter out relevant information. Acquiring high-quality data from this unfiltered dataset is crucial for building successful models. Additionally, the quantity and quality of data used for training the model plays a significant role. In essence, an ideal dataset should be specific to the targeted property, have sufficient quantity, and encompass various possible variations to have unbiased prediction. Unfortunately, obtaining an ideal dataset often becomes a bottleneck, which consequently affects the predictive capacity of the model, regardless of the effectiveness of the ML algorithm. Inconsistencies in the data can also undermine the model's performance. In the context of materials science, researchers are continuously exploring new dataset repositories and employing new methods of feature importance analysis. Predicting a potential material with targeted property from a wide chemical space holds particular importance in green energy technologies and other applications. To achieve accurate predictions, several factors come into play, including the choice of ML algorithm, dataset quality, feature selection, and model validation. The dataset should be representative, clean, errorfree, and contain relevant features. Understanding basic physical, chemical, and thermodynamical properties can help narrow down the search for essential features and guide the selection of ML mod

FAQs

What is the vacancy-ordered 18 valence electron (VHH) strategy in this study?

The work introduces ordered vacancies into half Heusler systems to convert a 19 VEC framework into an 18 VEC configuration, targeting semiconducting behavior and reduced thermal conductivity.

How is formation energy (ΔHf) predicted and validated?

A machine learning model is trained on 4684 Materials Project entries to predict ΔHf, then DFT calculations validate predictions for Zr0.75NiSb and Hf0.75NiSb, showing comparable ΔHf values.

What properties indicate the materials’ suitability for thermoelectrics?

Calculated electronic and lattice dynamics show narrow band-gap semiconductors and dynamical stability from positive phonon frequencies, while Seebeck coefficient, electrical/thermal conductivity, power factor, and ZT are evaluated.

Aug 6, 2026

Upload by Himbo