Uncertainty quantification reveals the limits of lattice thermal conductivity prediction
Accurate prediction of lattice thermal conductivity (κ L ) drives thermoelectric materials discovery, yet existing machine learning models rarely quantify the reliability of their predictions. Here, we use evidential neural networks to quantify predictive uncertainty and split it into aleatoric uncertainty (data noise) and epistemic uncertainty (model ignorance). Our analysis reveals that predictions of very low lattice thermal conductivity are inherently unreliable as aleatoric noise dominates the uncertainty, and this noise can rival or even exceed the predicted value itself. High-κ L materials, by contrast, carry higher epistemic uncertainty, indicating room for improvement by additional training data. Consistent with this observation, adding out-of-distribution samples improves model accuracy by reducing model uncertainty. We further develope an uncertainty-aware screening method that prioritizes predictions with low total uncertainty rather than relying solely on the absolute value of κ L . This approach recovers a wide range of compositional classes with low κL from the Starrydata experimental repository at nearly twice the rate of random selection, thereby narrowing the search space needed to identify promising materials. These results establish evidential neural networks as a practical uncertainty-aware tool for κ L prediction, enabling reliable candidate screening and guiding data-collection to accelerate materials discovery.
Authors
- Milad Rayka (ORCID: https://orcid.org/0000-0002-4378-9609)
- S. Shahab Naghavi (ORCID: https://orcid.org/0000-0002-8511-2591)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-19
- DOI
- https://doi.org/10.1038/s41598-026-81477-y
- Primary Topic
- Machine Learning in Materials Science
- Type
- article
- Field-Weighted Citation Impact
- 0.00