🔬 materials science

High-throughput screening of Heusler alloys for thermoelectrics combining machine-learned densities of states and first-principles transport

This study validates a high-throughput screening workflow for Heusler thermoelectrics that combines machine-learned density of states with first-principles transport calculations to efficiently identify 80 promising computational candidates, including 61 unreported materials, while demonstrating that although neural network predictions often miss band gaps, they remain effective for prioritizing compounds for detailed analysis.

Original authors: Shivam Randive, Rahul Dixit, Himanshu Pandey

Published 2026-10-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Shivam Randive, Rahul Dixit, Himanshu Pandey

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The world generates a staggering amount of waste heat, from the exhaust of cars to the cooling systems of power plants. This energy simply dissipates into the air, lost forever. Thermoelectric materials offer a way to capture that lost heat and turn it directly into electricity, acting as a silent generator that requires no moving parts. The efficiency of these materials depends on a delicate balance: they must conduct electricity well while blocking the flow of heat. Finding new materials that strike this balance is like searching for a needle in a haystack, because the standard computer simulations used to predict their performance are incredibly slow and expensive. Scientists often have to choose which of the millions of possible chemical combinations to test, but without a reliable way to predict the outcome, they risk missing the best candidates or wasting time on dead ends.

A team of researchers set out to solve this problem by testing a new strategy on a massive collection of known chemical structures called Heusler alloys. These are a specific family of compounds made from three different elements, known for their potential in energy applications. The researchers wanted to see if they could use a shortcut: a machine-learning model that predicts the electronic "fingerprint" of a material, known as the density of states, to decide which materials deserve a full, expensive computer simulation. They tested this approach on over one hundred thousand different Heusler structures. The goal was to see if the machine learning predictions were good enough to guide the search, and if the shortcut indicators they used could actually find the materials that would work as thermoelectrics.

The researchers began by dividing their massive database into two groups. The first group contained structures where the electronic fingerprint was already known from previous, high-quality calculations. The second group contained structures where this fingerprint was missing. For the second group, they trained a sophisticated computer program to predict the missing fingerprints based on the arrangement of atoms. They then applied a screening process to both groups. This process looked for specific signs in the electronic fingerprint that usually indicate a material might be a good thermoelectric. One sign was a simple check to see if there was a gap in the electron energy levels, which is necessary for the material to act as a semiconductor. A second, more sophisticated sign measured how the electrons would respond to heat, a property known as the Seebeck effect, which drives the generation of electricity.

The results revealed a surprising truth about the shortcuts. The simple check for an energy gap, which many researchers rely on, failed to do its job. Even when looking at the high-quality, known data, this simple check incorrectly labeled many metals as semiconductors. It was too easily fooled by small dips in the data that looked like gaps but were not. The machine learning model also struggled with this specific task; while it could reproduce the general shape of the electronic fingerprint with high accuracy, it frequently smoothed over the tiny gaps, making them disappear. This meant that the simple "gap or no gap" question could not be trusted, even when answered by the machine learning model.

However, the more sophisticated indicator, which measured the potential for generating voltage from heat, worked remarkably well. This indicator looked at the overall shape of the electronic fingerprint rather than just a single point. It successfully identified the promising materials, placing the best candidates at the very top of the list. When the researchers ran the full, expensive simulations on the top-ranked materials, they found that this indicator had indeed concentrated the winners. In the group of materials where the fingerprints were predicted by the machine learning model, nearly all the materials that passed the final tests were found within the top twenty percent of the list. This suggests that even though the machine learning model made mistakes about the existence of gaps, its overall prediction was good enough to guide the search effectively.

The study ultimately identified eighty distinct chemical compositions that are strong candidates for thermoelectric applications. However, it is crucial to note that sixty-one of these eighty compositions have no reported thermoelectric measurements for the Heusler phase; for these, the potential is based entirely on computational predictions. The researchers found that ten of the eleven best-known thermoelectric materials in this family were correctly identified by their method, proving the approach works for known chemistry. They also discovered that for materials containing very heavy elements, a subtle physical effect called spin-orbit coupling significantly changes the results, often reducing their performance. This effect had to be accounted for in the final calculations to avoid false positives. Furthermore, the study showed that changing the shape of the crystal structure slightly, such as stretching it from a cube to a box, rarely turned a metal into a semiconductor, meaning that the basic chemical composition is the most critical factor.

The researchers concluded that while machine learning cannot yet perfectly predict the existence of tiny energy gaps, it is still a powerful tool for sorting through vast databases. The key is to use the right kind of indicator. A simple yes-or-no check for a gap is unreliable, but a measure of the material's potential to generate voltage from heat is robust. By using this better indicator, scientists can drastically reduce the number of expensive simulations needed to find new materials. The eighty candidates identified in this study are now ready for the next stage of research, where scientists will need to verify their stability against competing crystal structures and measure their actual performance in the lab. This work provides a clear roadmap for how to use artificial intelligence to accelerate the discovery of materials that could one day help us harvest waste heat more efficiently.

Technical Summary: High-Throughput Screening of Heusler Alloys for Thermoelectrics

Problem Statement
The discovery of new thermoelectric materials is hindered by the computational cost of first-principles transport calculations, which are required to determine the figure of merit ( $zT$ ). While machine learning (ML) has enabled the prediction of electronic densities of states (DOS) for large databases, it remains untested whether DOS-derived indicators can reliably select compounds for expensive transport calculations, particularly when the DOS itself is predicted rather than calculated. Furthermore, existing high-throughput screens often focus on cubic structures, leaving tetragonal and metastable variants of Heusler compounds largely unexamined. This study addresses the question of whether a DOS-based screen, utilizing both Density Functional Theory (DFT) and ML-predicted spectra, can effectively identify thermoelectric candidates within the massive DXMag Heusler database (106,235 entries).

Methodology
The study employs a two-population screening strategy on the DXMag database, which contains 106,235 energy minima derived from 27,864 composition-prototype pairs.

Data Partitioning: The database is split into a training set (51,257 entries) with existing DFT spin- and orbital-resolved d-DOS, and a prediction set (54,978 entries) lacking DFT DOS.
Machine Learning Model: An Atomistic Line Graph Neural Network (ALIGNN) is trained on the training set to predict the spin-resolved d-DOS. The model uses a loss function heavily weighted near the Fermi level ( $E_F$ ) to prioritize thermoelectric-relevant features.
Screening Indicators: Two DOS-derived descriptors are tested:
- Binary Gap Label: A threshold-based check for a gap in the d-DOS at $E_F$ .
- Seebeck Indicator ( $S^d_{max}$ ): A metric derived from the Cutler–Mott approximation, calculating the maximum thermopower achievable by shifting the chemical potential ( $\mu$ ) within $\pm 1$ eV of $E_F$ , weighted by the thermal window at 900 K.
Selection Protocol: Entries are filtered by stability (convex hull distance), electron counting (18 or 24 valence electrons), and the DOS indicators. Selected entries undergo rigorous first-principles assessment using Quantum ESPRESSO (QE-PBE) for band structures and BoltzTraP2 for transport coefficients under the constant-relaxation-time approximation (CRTA).
Certification Criteria: A compound is "certified" if it satisfies five sequential criteria: (i) $|S| > 60$ $\mu$ V K $^{-1}$ , (ii) Power Factor/relaxation time ( $PF/\tau$ ) $> 1.09 \times 10^{12}$ W m $^{-1}$ K $^{-2}$ s $^{-1}$ , (iii) thermodynamic stability (hull distance $< 0.10$ eV atom $^{-1}$ ), (iv) a QE-PBE band gap $\ge 0.136$ eV, and (v) for heavy elements ( $Z \ge 72$ ), validation with spin–orbit coupling (SOC).

Key Results

Performance of ML-Predicted Spectra: The ALIGNN model reproduces the general shape of the d-DOS with high correlation ( $r \approx 0.895$ ) but fails to accurately predict band gaps. The recall for detecting gaps in held-out spectra is only 15%, as the model tends to smooth over gaps with residual weight. However, the spectral similarity metrics (like correlation) do not capture this failure.
Failure of Binary Gap Labels: Even when using accurate DFT spectra, the binary gap label is a poor classifier. It identifies many metals as semiconductors because it ignores s- and p-states within a d-orbital gap and relies on a fixed threshold. Most entries passing the label are ultimately metallic in band-structure calculations.
Success of the Seebeck Indicator: The $S^d_{max}$ $S_{ma x}^{d}$ indicator successfully concentrates certified entries at the top of the ranking.
- In the training set, the top 20% of entries ranked by $S^d_{max}$ contain 75% of all certified entries.
- In the prediction set, the top 20% contain 91% of certified entries.
- Crucially, $S^d_{max}$ remains effective even on ML-predicted spectra, despite the model's inability to resolve gaps perfectly. The indicator's integration over a thermal window makes it tolerant of the residual weight inside predicted gaps.
Role of Spin-Orbit Coupling (SOC): For heavy-element candidates, SOC significantly suppresses p-type power factors (median ratio of 0.76 compared to scalar-relativistic values) due to the splitting of degenerate valence-band states, while n-type factors remain largely unaffected. This effect eliminates 40 of 115 heavy-element candidates that were initially qualified.
Structural Variants: Screening tetragonal and metastable structural variants revealed that they rarely change the electronic class of a composition. No variant was found to be gapped where its reference (lowest-energy) crystal was metallic.
Certified Candidates: The screen identified 80 distinct compositions (155 entries) that passed all criteria.
- Validation: Ten of the eleven established half-Heusler thermoelectrics (e.g., TiNiSn, ZrNiSn, NbFeSb) were successfully certified, validating the protocol.
- Novelty: 61 of the 80 compositions have no reported thermoelectric measurements for the Heusler phase.
- Phase Stability: Several certified compositions (e.g., those with Si, Ge, or As on the Z-site) are flagged because experimental or predicted orthorhombic phases (TiNiSi-type) compete with the half-Heusler structure, a limitation of the current convex hull approach.

Significance and Claims
The paper claims that a DOS-based screen can effectively prioritize compounds for transport calculations, but it emphasizes specific limitations and insights:

Indicator vs. Ranking: The DOS-derived Seebeck indicator is a powerful tool for prioritizing candidates (concentrating them at the top of the list) but cannot rank them by power factor. The power factor depends on band velocities and valley degeneracy, which the DOS alone does not capture.
Necessity of Band Structure: Neither a binary gap label nor an ML-predicted spectrum can definitively confirm the existence of a band gap. A full band-structure calculation is required to certify a gap.
Utility of ML: While ML-predicted spectra fail to reproduce gaps accurately, the screening workflow remains robust because the Seebeck indicator is tolerant of the specific errors (residual weight) introduced by the model.
Scope of Candidates: The 80 certified compositions are presented as computational candidates for further investigation (beyond CRTA, including scattering and lattice thermal conductivity) and synthesis, rather than as confirmed high-performance thermoelectrics. The study explicitly notes that the list is a prioritized sample, not a census, and is biased against wide-gap compounds due to the chemical potential window used in certification.

In conclusion, the study demonstrates that combining ML-predicted DOS with a physics-informed Seebeck indicator allows for the efficient screening of massive databases, successfully recovering known thermoelectrics and identifying new candidates, provided that final certification relies on first-principles band structures and transport calculations.

Technical Summary: High-Throughput Screening of Heusler Alloys for Thermoelectrics

More like this