Beyond Predicted ZT: Machine Learning Strategies for the Experimental Discovery of Thermoelectric Materials

TL;DR

This paper reviews machine learning strategies for discovering thermoelectric materials, highlighting challenges like data limitations and validation issues, and proposes advanced methods to improve experimental success rates.

Contribution

It identifies key obstacles in ML-based thermoelectric discovery and advocates for innovative validation and active learning approaches to enhance experimental validation.

Findings

ML models often overestimate performance due to sampling biases

Poor model generalizability limits discovery of high-zT materials

Active learning strategies can bridge the gap between prediction and experiment

Abstract

The discovery of high-performance thermoelectric (TE) materials for advancing green energy harvesting from waste heat is an urgent need in the context of looming energy crisis and climate change. The rapid advancement of machine learning (ML) has accelerated the design of thermoelectric (TE) materials, yet a persistent "gap" remains between high-accuracy computational predictions and their successful experimental validation. While ML models frequently report impressive test scores (R^2 values of 0.90-0.98) for complex TE properties (zT, power factor, and electrical/thermal conductivity), only a handful of these predictions have culminated in the experimental discovery of new high-zT materials. In this review, we identify and discuss that the primary obstacles are poor model generalizability-stemming from the "small-data" problem, sampling biases in cross-validation, and inadequate structural representation-alongside the critical challenge of thermodynamic phase stability. Moreover, we argue that standard randomized validation often overestimates model performance by ignoring "hidden hierarchies" and clustering within chemical families. Finally, to bridge this gap between ML-predictions and experimental realization, we advocate for advanced validation strategies like PCA-based sampling and a synergetic active learning loop that integrates ML "fast filters" for stability (e.g., GNoME) with high-throughput combinatorial thin-film synthesis to rapidly map stable, high-zT compositional spaces.