Unsupervised machine learning for discovery of promising half-Heusler thermoelectric materials

Abstract

Thermoelectric materials can be potentially applied to waste heat recovery and solid-state cooling because they allow a direct energy conversion between heat and electricity and vice versa. The accelerated materials design based on machine learning has enabled the systematic discovery of promising materials. Herein we proposed a successful strategy to discover and design a series of promising half-Heusler thermoelectric materials through the iterative combination of unsupervised machine learning with the labeled known half-Heusler thermoelectric materials. Subsequently, optimized zT values of ~0.5 at 925 K for p-type Sc0.7Y0.3NiSb0.97Sn0.03 and ~0.3 at 778 K for n-type Sc0.65Y0.3Ti0.05NiSb were experimentally achieved on the same parent ScNiSb.

Machine learned synthesizability predictions aided by density functional theory

“…Anand et al 47 calculated E hull to evaluate DFT stability while Vikram et al 48 used E f (formation energy) to assess DFT stability for identifying synthesizable half-Heuslers. Jia et al 49 more recently used a criteria of E hull < 100 meV/atom for synthesizability in addition to other clustering techniques to to identify new half-Heuslers candidates. Other works have found magnetic Heuslers 50 , low thermal conductivity Heuslers 51 , and others with various novel properties 2,[52][53][54][55][56] .…”

Rational ensemble design of alloy catalysts for selective ammonia oxidation based on machine learning

“…An unsupervised clustering model could quickly identify the narrow region of containing targeted materials to accelerate the discovery of new materials, based on the unlabelled data with only input features. 30,31,45 The K -means model that has exhibited excellent clustering performance on half-Heusler thermoelectric materials discovery 30 is adopted to identify promising ensemble configurations for NH 3 -SCO catalysts merely based on the features of the binding energy of intermediates, as shown in Fig. 6.…”

“…Liu et al proposed a multiple iterative clustering strategy of 456 half-Heusler materials data set with 20 known thermoelectric materials, and Sc 0.7 Y 0.3 NiSb 0.97 Sn 0.03 and Sc 0.65 Y 0.3 Ti 0.05 NiSb are obtained and experimentally validated as promising thermoelectric materials. 30 Ling et al utilized a small quantity of conductivity data to perform the clustering study and eventually discovered 16 new fast Li-conductors from a large amount of Li-containing materials. 31 These cases confirm the strong ability of the unsupervised machine learning model to discover some promising materials from a large unknown material space only based on some limited performance data.…”

A Critical Review of Machine Learning Techniques on Thermoelectric Materials

“…Screening combination of unsupervised machine learning with the labeled reported known half-Heusler TE materials. Reproduced with permission from ref . Copyright 2022 the authors.…”

“…Along with commonly used supervised ML algorithms, Jia et al proposed an unsupervised ML screening strategy to find entirely new and promising half-Heusler TE systems (Figure 8). 98 First, 456 different half-Heusler compounds were extracted from the Materials Project database and 20 reported half-Heusler TE compounds were labeled. Three different types of unsupervised clustering algorithms were employed to classify half-Heusler compounds.…”