Combining Machine Learning Models with First-Principles High-Throughput Calculation - Accelerate Thermoelectric Material Search

Aug 5, 2026

Upload by Asher

Thermoelectric materials enable direct electricity–heat conversion for waste-heat harvesting and solid-state cooling, but finding new candidates is often limited by expensive, time-consuming experimentation and first-principles transport calculations. The work replaces costly screening with a hybrid pipeline: large-scale first-principles high-throughput calculations for 796 chalcogenide compounds to build a thermoelectric database, followed by training six predictive models. Four ensemble and two deep-learning models classify promising n-type and p-type dopants, achieving classification accuracy above 85% and AUC above 0.9, with the M3GNet n-type model exceeding 90% across accuracy, precision, and recall.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages31

Reading Time1 h 18 min

AccessFree to download

Abstract

Introduction

Thermoelectric materials and performance metrics (ZT)

Traditional discovery methods and computational cost

Machine learning in thermoelectrics and materials science

Building a high-throughput thermoelectric database

Hybrid ML + first-principles screening strategy

Combining Machine Learning Models with First-Principles High-Throughput Calculation to Accelerate the Search of Promising Thermoelectric Materials Tao Fan*, Artem R. Oganov Skolkovo Institute of Science and Technology, Bolshoy Boulevard 30, bld. 1, 121205 Moscow, Russia. ABSTRACT Thermoelectric materials can achieve direct energy conversion between electricity and heat, thus can be applied to waste heat harvesting and solid-state cooling. The discovery of new thermoelectric materials is mainly based on experiments and first-principles calculations. However, these methods are usually expensive and time-consuming. Recently, the prediction of properties via machine learning has emerged as a popular method in materials science. Herein, we firstly did first-principles high-throughput calculations for a large number of chalcogenides and built a thermoelectric database containing 796 compounds. Many novel and promising thermoelectric materials were discovered. Then, we trained four ensemble learning models and two deep learning models to distinguish the promising thermoelectric materials from the others for n type and p type doping, respectively. All the presented models achieve classification accuracy higher than 85% and area under the curve (AUC) higher than 0.9. Especially, the M3GNet model for n type data achieve accuracy, precision and recall all higher than 90% . Our works demonstrate a very efficient way of combining machine learning prediction and first-principles high-throughput calculations together to accelerate the discovery of advanced thermoelectric materials. INTRODUCTION Thermoelectric (TE) materials could play an important role in building clean and alternative energy sources due to their ability to realize the direct conversion between heat and electricity [1-3] . Thermoelectric devices have the characteristics of small size, no noise, and no pollution, thus they have wide application in space power, industrial waste heat harvesting, small and mobile refrigerators, and other fields [4-6] . The energy conversion efficiency of thermoelectric materials depends on the dimensionless figure of merit (ZT) . ZT is defined as 􀜼􀜶 = 􀟙 2 􀟪􀜶⁄(􀟢􀯘 + 􀟢􀯅) , where α is the Seebeck coefficient, σ is the electrical conductivity, T is the absolute temperature, κe is the electronic thermal conductivity, and κL is the lattice thermal conductivity. Particularly,􀟙 2 􀟪 is called the power factor (PF) . In order to obtain a high ZT, both α and σ must be maximized, while κe and κL need to be minimized. However, the interdependence of these parameters makes improving the ZT ofa material a great challenge[7,8] . Traditional thermoelectric materials discovery has been led by experiments, while computations are becoming more and more important with the advance of theory and the increase of computing power [9-11] . First-principles methods, such as DFT, have been widely used in calculating the thermoelectric related properties [12-15] . However, full first-principles calculation of transport properties is usually computationally expensive. Thus, there are many simplified models being proposed to calculate electronic and phonon transport properties, leading to many interesting and important discoveries[16-20]. However, they all face the problem of accuracy-computational cost trade-off. Recently, machine learning (ML) has achieved much progress in both its theory and available models [21] . Data science and machine learning have become an integral part of natural sciences, thought as the fourth pillar in science, next to experiment, theory, and simulation [22-25] . ML algorithms find patterns in highdimensional training data and build a mathematical model to make predictions or decisions without explicit human knowledge. This approach has been applied successfully to various materials science studies, such as structure predictions [26,27], constructing force field [28-30], and predictions of the static properties of materials[31- 33

FAQs

Why are machine learning models helpful for discovering thermoelectric materials?

First-principles transport calculations are computationally expensive, while ML can learn from existing data to predict promising thermoelectric candidates quickly with reduced cost.

How is the thermoelectric dataset constructed in this work?

First-principles high-throughput calculations are performed for 796 chalcogenide compounds, producing n-type and p-type thermoelectric properties to form a dedicated database.

What performance do the trained models achieve for n-type and p-type classification?

The ensemble and deep learning models reach classification accuracy above 85% and AUC above 0.9, with the M3GNet n-type model achieving accuracy, precision, and recall all above 90%.

Aug 5, 2026

Upload by Asher