Thermoelectric ML Framework: Physical Consistency and ZT Screening

PLOS ONE（Dongguan University of Technology）

Overview

A physics-guided machine learning framework has been developed to screen thermoelectric materials by independently predicting transport properties rather than direct ZT regression. By integrating uncertainty propagation and material-wise validation, the approach reduces target leakage and provides more physically consistent performance estimations.

In Depth

Key Findings

The developed machine learning (ML) framework demonstrates a robust strategy for thermoelectric materials screening by prioritizing physical consistency over apparent predictive accuracy. Unlike direct regression models that predict the figure of merit (ZT), this framework independently predicts the Seebeck coefficient, electrical conductivity, and thermal conductivity. While this method shows lower apparent accuracy than direct ZT models, it provides improved interpretability and eliminates target leakage caused by the mathematical derivation of ZT from multiple transport properties.

Technical Details

The framework utilizes several specific methodologies to ensure reliability:

• Material-wise Validation: To prevent material-level information leakage, exact chemical formulas are kept exclusively in either the training or held-out test subsets, rather than distributing measurements from the same material across both.
• Uncertainty Quantification: Predictive uncertainty is quantified using ensemble learning. This uncertainty is then propagated into the thermoelectric performance evaluation and uncertainty-aware Lower Confidence Bound (LCB) screening.
• Data Source: The models were evaluated using experimental data from the ESTM database.

Background & Context

Thermoelectric materials convert heat into electricity via the Seebeck effect, making them valuable for waste heat recovery and localized cooling. However, discovery is hindered by the interdependence of the Seebeck coefficient (S), electrical conductivity (σ), and thermal conductivity (κ), where improving one often negatively impacts another. Historically, this has required resource-intensive experimental iteration. While supervised learning models like random forests and deep neural networks have been applied to accelerate this process, many suffer from overly optimistic performance due to hidden statistical correlations in the datasets.

Significance & Outlook

By integrating physical consistency and uncertainty propagation, this framework offers a more realistic approach to identifying candidate materials. The transition from direct ZT prediction to a structured, property-based evaluation allows researchers to avoid the pitfalls of target leakage and provides a more reliable roadmap for data-driven thermoelectric materials discovery.

Get our weekly technology intelligence — free

Receive an infographic that lets you judge at a glance whether each field’s analysis report is worth reading.

Subscribe Free — Weekly Tech Intelligence

By subscribing, you’ll receive Troy-Technical’s weekly technology intelligence newsletter.

• Your email and selected fields are used only to deliver the newsletter.
• We never share your information with third parties.
• You can unsubscribe anytime via the link in each email.

See our Privacy Policy for details.

Agree & Continue ▶

Takes about a minute · Unsubscribe anytime