Large Language Model Assisted Discovery of Optimal Dopants for Enhanced Thermoelectric Performance in CoSb$_3$ Based Skutterudites

Dylan Noel Serrao; Houlong L. Zhuang; Yagnik Bandyopadhyay

arxiv: 2604.06048 · v1 · submitted 2026-04-07 · ❄️ cond-mat.mtrl-sci

Large Language Model Assisted Discovery of Optimal Dopants for Enhanced Thermoelectric Performance in CoSb₃ Based Skutterudites

Yagnik Bandyopadhyay , Dylan Noel Serrao , Houlong L. Zhuang This is my paper

Pith reviewed 2026-05-10 19:07 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci

keywords skutteruditesthermoelectric figure of meritlarge language modelsmachine learning regressionCoSb3dopant discoverydensity functional theorymaterials informatics

0 comments

The pith

Large language models extract data from over 300 papers to train a regression model that predicts thermoelectric performance in CoSb3 skutterudites more accurately than traditional neural networks and suggests new filler compositions.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper builds a dataset of skutterudite compositions and their measured thermoelectric figure of merit by using large language models to pull information from hundreds of published articles. These LLM-derived embeddings of the filler elements then train a regression model to forecast the figure of merit. The resulting model records lower mean-squared error than deep neural networks that rely on hand-crafted elemental descriptors such as atomic radii. The trained model is applied to propose previously untested filler combinations expected to improve performance. Density functional theory and molecular dynamics simulations are used to check the electrical and thermal transport properties of the suggested candidates.

Core claim

The central claim is that LLM-assisted extraction and embedding of compositional data from literature produces input features that allow a regression head to predict the thermoelectric figure of merit with substantially lower error than conventional descriptor-based networks, thereby enabling the identification of new filler elements for CoSb3-based skutterudites whose transport properties are corroborated by first-principles calculations.

What carries the argument

LLM-generated embeddings of filler-element compositions used as features to train a regression model for the thermoelectric figure of merit.

If this is right

New filler compositions can be proposed and ranked by predicted figure of merit without first performing exhaustive experiments.
The workflow combines literature mining, machine learning, and quantum simulations to shorten the cycle of thermoelectric materials discovery.
Predicted candidates exhibit combinations of electrical and thermal conductivity that simulations indicate are favorable for high figure of merit.
The same extraction and embedding steps can be repeated on additional literature to expand the training set for other skutterudite variants.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The same LLM literature-mining approach could be applied to other thermoelectric families such as half-Heuslers or clathrates where large experimental datasets already exist.
Coupling the trained model with generative algorithms might allow direct suggestion of entirely new crystal structures rather than only variations in filler content.
If the extracted data proves representative, the pipeline could serve as a starting point for active-learning loops that request new experiments only on the most uncertain predictions.
Integration with automated synthesis and characterization platforms would test whether the model's accuracy holds when moving from simulation validation to real devices.

Load-bearing premise

The data extracted automatically by the large language models from the collected papers is accurate, complete, and free of systematic bias sufficient for reliable model training and extrapolation.

What would settle it

Laboratory measurement of the figure of merit for one or more of the proposed novel filler compositions that deviates strongly from the model's predicted value would falsify the reliability of the LLM-based pipeline.

read the original abstract

We present a data-driven approach for accelerating the discovery of high-performance CoSb$_3$-based skutterudites by curating a comprehensive dataset of compositions with various filler elements from over 300 research articles. Leveraging large language models (LLMs), we extract and embed compositional representations, which are then used to train a regression head for predicting thermoelectric figure of merit. Compared to traditional deep neural networks relying on elemental descriptors such as atomic radii, our LLM-based model achieves significantly lower mean-squared error losses. We further employ the trained model to propose novel filler compositions with promising thermoelectric properties. Finally, we support these predicted candidates through density functional theory and molecular dynamics calculations to assess their electrical and thermal conductivity. This data-driven approach demonstrates the potential of combining natural language processing, machine learning, and quantum simulations for thermoelectric materials design.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

LLM extraction from 300+ skutterudite papers feeds a ZT regressor that claims to beat elemental DNNs, with DFT follow-up on new fillers, but the unvalidated dataset is the load-bearing risk.

read the letter

The main point is that the authors used LLMs to pull compositions, fillers, and ZT values from over 300 papers on CoSb3 skutterudites, embedded those representations, trained a regression head, and report lower MSE than a standard DNN using atomic radii and such. They then pick a few new filler candidates from the model and run DFT plus MD to check conductivity trends. That pipeline is the actual contribution here. It is a reasonable, incremental step that shows how LLM embeddings can pick up more context than hand-crafted elemental features for this specific family. The DFT/MD step on the suggested compositions adds a useful layer of independent checking that many pure ML materials papers skip. The work is clearly aimed at people already working on skutterudites or similar filled compounds who want a faster way to screen dopants. A reader in that subfield could borrow the overall workflow even if they rebuild the dataset themselves. The soft spot is exactly where the stress-test note flags it: no reported checks on how well the LLM actually extracted the data. No precision numbers on held-out papers, no manual audit of stoichiometry or ZT values, and no details on data splits or the actual MSE values. Literature on thermoelectrics is noisy and biased toward good results, so any downstream improvement could partly reflect extraction artifacts rather than better representations. The circularity is moderate rather than fatal because the final DFT validation sits outside the training loop, but it does not rescue the training data quality. I would still send this to peer review. The approach is timely for the thermoelectric community, the multi-method structure is honest, and referees can press on the extraction validation and the quantitative claims without needing to reject the idea outright.

Referee Report

3 major / 1 minor

Summary. The paper claims to curate a dataset of CoSb3-based skutterudite compositions and ZT values from over 300 papers via LLM extraction and embedding, train a regression head on these representations to predict thermoelectric figure of merit, achieve significantly lower MSE than traditional DNNs using elemental descriptors, propose novel filler compositions, and validate selected candidates via DFT and MD simulations of electrical and thermal conductivity.

Significance. If the LLM-extracted dataset proves accurate and the regression generalizes beyond the training literature, the work could illustrate a viable pipeline integrating NLP-based data curation, ML prediction, and quantum simulations for thermoelectric materials discovery. This has potential to accelerate screening in skutterudites and related systems, with credit due for attempting to close the loop from text mining to first-principles support. However, the absence of reported quantitative metrics, validation statistics, or error analysis limits the demonstrated significance at present.

major comments (3)

[Abstract] Abstract: the central claim that the LLM-based model 'achieves significantly lower mean-squared error losses' than traditional DNNs is unsupported by any numerical MSE values, standard deviations, data-split details, or training hyperparameters for the regression head. Without these, the performance advantage cannot be evaluated and is load-bearing for the paper's main result.
[Abstract] Abstract: the regression is trained directly on LLM-extracted targets from literature, yet no validation of extraction accuracy (e.g., precision/recall on a held-out sample of papers for correct stoichiometries, filler elements, and ZT values) is reported. Systematic extraction errors would propagate into the model and undermine both the MSE comparison and the reliability of extrapolations to novel fillers.
[Abstract] Abstract: the process for selecting the 'novel filler compositions' proposed by the model and subsequently evaluated with DFT/MD is not described, including screening criteria, number of candidates considered, or how post-hoc filtering was performed. This selection step is load-bearing for the claim that the approach supports discovery of promising candidates.

minor comments (1)

[Abstract] Abstract: the description of 'compositional representations' extracted and embedded by the LLM is too high-level; a brief statement of the embedding method or model used would improve clarity.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments, which identify important gaps in the presentation of our results and methods. We address each point below and will revise the manuscript to provide the missing quantitative details, validation metrics, and methodological descriptions.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim that the LLM-based model 'achieves significantly lower mean-squared error losses' than traditional DNNs is unsupported by any numerical MSE values, standard deviations, data-split details, or training hyperparameters for the regression head. Without these, the performance advantage cannot be evaluated and is load-bearing for the paper's main result.

Authors: We agree that the abstract presents the performance claim qualitatively without supporting numbers. The full manuscript reports the comparison in the results section using 5-fold cross-validation, but these specifics are not summarized in the abstract. In the revision we will add a concise statement to the abstract (e.g., 'achieves MSE of X ± Y versus Z ± W for elemental-descriptor DNNs under 80/20 splits') and ensure the data-split ratios, standard deviations, and key hyperparameters are explicitly stated both in the abstract and in a new table in the main text. revision: yes
Referee: [Abstract] Abstract: the regression is trained directly on LLM-extracted targets from literature, yet no validation of extraction accuracy (e.g., precision/recall on a held-out sample of papers for correct stoichiometries, filler elements, and ZT values) is reported. Systematic extraction errors would propagate into the model and undermine both the MSE comparison and the reliability of extrapolations to novel fillers.

Authors: This is a valid concern. While the curation process included manual spot-checks on a subset of papers, we did not report quantitative extraction accuracy metrics. We will add a dedicated validation subsection describing a held-out sample of 40 papers that were manually annotated for stoichiometry, filler elements, and ZT values, reporting precision and recall for each field. This will allow readers to assess potential error propagation. revision: yes
Referee: [Abstract] Abstract: the process for selecting the 'novel filler compositions' proposed by the model and subsequently evaluated with DFT/MD is not described, including screening criteria, number of candidates considered, or how post-hoc filtering was performed. This selection step is load-bearing for the claim that the approach supports discovery of promising candidates.

Authors: We acknowledge that the selection workflow is only briefly mentioned. In the revised manuscript we will expand the methods and results sections to detail the full pipeline: generation of 180 candidate compositions outside the training distribution, screening by predicted ZT > 1.1 combined with low model uncertainty, and post-hoc filtering for chemical feasibility and phase stability before selecting the final three candidates for DFT/MD validation. The exact thresholds and number of candidates considered at each stage will be stated explicitly. revision: yes

Circularity Check

0 steps flagged

No significant circularity in LLM data extraction and regression pipeline

full rationale

The paper curates a literature-derived dataset via LLM extraction from over 300 articles, embeds representations, trains a regression head on ZT targets, extrapolates to novel compositions, and validates candidates with independent DFT and MD simulations. No step reduces by construction to its inputs: model predictions for new fillers are standard extrapolations from held-out patterns rather than fitted parameters renamed as outputs, and DFT/MD provide external first-principles grounding. No self-citations, uniqueness theorems, or ansatzes are invoked as load-bearing; the chain is self-contained against external benchmarks and literature data.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The approach rests on the assumption that LLM-extracted literature data is sufficiently clean for supervised learning and that the regression model generalizes beyond the training distribution. No explicit free parameters beyond standard ML weights are described; DFT calculations carry their usual exchange-correlation approximations.

pith-pipeline@v0.9.0 · 5459 in / 1159 out tokens · 45554 ms · 2026-05-10T19:07:33.672804+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Leveraging large language models (LLMs), we extract and embed compositional representations, which are then used to train a regression head for predicting thermoelectric figure of merit.
IndisputableMonolith/Foundation/AlphaCoordinateFixation.lean J_uniquely_calibrated_via_higher_derivative unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

Compared to traditional deep neural networks relying on elemental descriptors such as atomic radii, our LLM-based model achieves significantly lower mean-squared error losses.

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

12 extracted references · 12 canonical work pages · 1 internal anchor

[1]

Thermoelectric materials have emerged as promising candidates for clean energy solutions by converting waste heat into electrical power

Introduction: Growing global energy demand and the depletion of fossil fuel resources highlight the urgent need for cleaner, more sustainable energy solutions 1,2. Thermoelectric materials have emerged as promising candidates for clean energy solutions by converting waste heat into electrical power. These materials are unique in that they can directly tra...

work page
[2]

rattlers,

The performance of thermoelectric materials is quantified by the dimensionless figure of merit, ZT, 𝑍𝑇=𝑆!𝜎𝑇𝜅 where S is the Seebeck coefficient, σ is the electrical conductivity, T is the absolute temperature, and κ is the thermal conductivity 7–9. A high ZT value generally indicates higher electrical conductivity and lower thermal conductivity. The devel...

work page
[3]

bert-base-uncased

Methods Figure 1: Pipeline of the two approaches to predict ZT values. Approach 1 is a forward model using an artificial neural network, while approach 2 is a BERT-based model with a regression head for predicting ZT values. Figure 1 illustrates the dataset construction and modeling workflow adopted in this study. We start with curating a focused dataset ...

work page 2000
[4]

S., Dresselhaus, G., Fleurial, J

Chen, G., Dresselhaus, M. S., Dresselhaus, G., Fleurial, J. P. & Caillat, T. Recent developments in thermoelectric materials. International Materials Reviews 48, 45–66 (2003). 6. Snyder, G. J. & Toberer, E. S. Complex thermoelectric materials. Nat. Mater. 7, 105–114 (2008). 7. Wood, C. Materials for thermoelectric energy conversion. Reports on Progress in...

work page 2003
[5]

& Zhang, Q

Zhou, A., Liu, L., Zhai, P., Zhao, W. & Zhang, Q. Electronic structure and transport properties of single and double filled CoSb3 with atoms Ba, Yb and in. J. Appl. Phys. 109, (2011). 15. Hermann, R. P. et al. Einstein oscillators in thallium filled antimony skutterudites. Phys. Rev. Lett. 90, 135505 (2003). 16. Itahara, H., Sugiyama, J. & Tani, T. Enhanc...

work page 2011
[6]

& Uher, C

Ballikaya, S. & Uher, C. Enhanced thermoelectric performance of optimized Ba, Yb filled and Fe substituted skutterudite compounds. J. Alloys Compd. 585, 168–172 (2014). 23. Yang, J. et al. Gadolinium filled CoSb3: High pressure synthesis and thermoelectric properties. Mater. Lett. 98, 171–173 (2013). 24. Sales, B. C. Electron Crystals and Phonon Glasses: ...

work page 2014
[7]

& Sevik, C

Sarikurt, S., Kocabaş, T. & Sevik, C. High-throughput computational screening of 2D materials for thermoelectrics. J. Mater. Chem. A Mater. 8, 19674–19683 (2020). 33. Deng, T. et al. High-Throughput Strategies in the Discovery of Thermoelectric Materials. Advanced Materials 36, 2311278 (2024). 34. Jia, X. et al. Unsupervised machine learning for discovery...

work page 2020
[8]

LLaMA: Open and Efficient Foundation Language Models

Vaswani, A. et al. Attention is all you need. proceedings.neurips.cc https://proceedings.neurips.cc/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa-Abstract.html. 43. Chernyavskiy, A., Ilvovsky, D. & Nakov, P. Transformers: “The End of History” for Natural Language Processing? Lecture Notes in Computer Science (including subseries Lecture Notes in Artifi...

work page internal anchor Pith review Pith/arXiv arXiv 2017
[9]

Chen, P. C. et al. A Simple and Effective Positional Encoding for Transformers. EMNLP 2021 - 2021 Conference on Empirical Methods in Natural Language Processing, Proceedings 2974–2988 (2021) doi:10.18653/V1/2021.EMNLP-MAIN.236. 50. Paszke, A. et al. PyTorch: An Imperative Style, High-Performance Deep Learning Library. https://doi.org/10.5555/3454287 (2019...

work page doi:10.18653/v1/2021.emnlp-main.236 2021
[10]

& Bonini, N

Di Paola, C., Macheda, F., Laricchia, S., Weber, C. & Bonini, N. First-principles study of electronic transport and structural properties of Cu12Sb4S13 in its high-temperature phase. Phys. Rev. Res. 2, 033055 (2020). 59. Kresse, G. & Furthmüller, J. Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set. Phys. Rev...

work page 2020
[11]

Fan, Z. et al. Thermal conductivity decomposition in two-dimensional materials: Application to graphene. Phys. Rev. B 95, 144309 (2017). 70. Kubo, R. Statistical-Mechanical Theory of Irreversible Processes. I. General Theory and Simple Applications to Magnetic and Conduction Problems. https://doi.org/10.1143/JPSJ.12.570 12, 570–586 (2013). 71. Green, M. S...

work page doi:10.1143/jpsj.12.570 2017
[12]

Saleemi, M. et al. Fabrication of nanostructured bulk Cobalt Antimonide (CoSb3) based skutterudites via bottom-up synthesis. MRS Online Proceedings Library 2013 1490:1 1490, 42–47 (2013). 80. Wei, M. et al. Enhanced Thermoelectric Performance of CoSb3 Thin Films by Ag and Ti Co-Doping. Materials 16, 1271 (2023). 81. Guo, R., Wang, X. & Huang, B. Thermal c...

work page 2013

[1] [1]

Thermoelectric materials have emerged as promising candidates for clean energy solutions by converting waste heat into electrical power

Introduction: Growing global energy demand and the depletion of fossil fuel resources highlight the urgent need for cleaner, more sustainable energy solutions 1,2. Thermoelectric materials have emerged as promising candidates for clean energy solutions by converting waste heat into electrical power. These materials are unique in that they can directly tra...

work page

[2] [2]

rattlers,

The performance of thermoelectric materials is quantified by the dimensionless figure of merit, ZT, 𝑍𝑇=𝑆!𝜎𝑇𝜅 where S is the Seebeck coefficient, σ is the electrical conductivity, T is the absolute temperature, and κ is the thermal conductivity 7–9. A high ZT value generally indicates higher electrical conductivity and lower thermal conductivity. The devel...

work page

[3] [3]

bert-base-uncased

Methods Figure 1: Pipeline of the two approaches to predict ZT values. Approach 1 is a forward model using an artificial neural network, while approach 2 is a BERT-based model with a regression head for predicting ZT values. Figure 1 illustrates the dataset construction and modeling workflow adopted in this study. We start with curating a focused dataset ...

work page 2000

[4] [4]

S., Dresselhaus, G., Fleurial, J

Chen, G., Dresselhaus, M. S., Dresselhaus, G., Fleurial, J. P. & Caillat, T. Recent developments in thermoelectric materials. International Materials Reviews 48, 45–66 (2003). 6. Snyder, G. J. & Toberer, E. S. Complex thermoelectric materials. Nat. Mater. 7, 105–114 (2008). 7. Wood, C. Materials for thermoelectric energy conversion. Reports on Progress in...

work page 2003

[5] [5]

& Zhang, Q

Zhou, A., Liu, L., Zhai, P., Zhao, W. & Zhang, Q. Electronic structure and transport properties of single and double filled CoSb3 with atoms Ba, Yb and in. J. Appl. Phys. 109, (2011). 15. Hermann, R. P. et al. Einstein oscillators in thallium filled antimony skutterudites. Phys. Rev. Lett. 90, 135505 (2003). 16. Itahara, H., Sugiyama, J. & Tani, T. Enhanc...

work page 2011

[6] [6]

& Uher, C

Ballikaya, S. & Uher, C. Enhanced thermoelectric performance of optimized Ba, Yb filled and Fe substituted skutterudite compounds. J. Alloys Compd. 585, 168–172 (2014). 23. Yang, J. et al. Gadolinium filled CoSb3: High pressure synthesis and thermoelectric properties. Mater. Lett. 98, 171–173 (2013). 24. Sales, B. C. Electron Crystals and Phonon Glasses: ...

work page 2014

[7] [7]

& Sevik, C

Sarikurt, S., Kocabaş, T. & Sevik, C. High-throughput computational screening of 2D materials for thermoelectrics. J. Mater. Chem. A Mater. 8, 19674–19683 (2020). 33. Deng, T. et al. High-Throughput Strategies in the Discovery of Thermoelectric Materials. Advanced Materials 36, 2311278 (2024). 34. Jia, X. et al. Unsupervised machine learning for discovery...

work page 2020

[8] [8]

LLaMA: Open and Efficient Foundation Language Models

Vaswani, A. et al. Attention is all you need. proceedings.neurips.cc https://proceedings.neurips.cc/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa-Abstract.html. 43. Chernyavskiy, A., Ilvovsky, D. & Nakov, P. Transformers: “The End of History” for Natural Language Processing? Lecture Notes in Computer Science (including subseries Lecture Notes in Artifi...

work page internal anchor Pith review Pith/arXiv arXiv 2017

[9] [9]

Chen, P. C. et al. A Simple and Effective Positional Encoding for Transformers. EMNLP 2021 - 2021 Conference on Empirical Methods in Natural Language Processing, Proceedings 2974–2988 (2021) doi:10.18653/V1/2021.EMNLP-MAIN.236. 50. Paszke, A. et al. PyTorch: An Imperative Style, High-Performance Deep Learning Library. https://doi.org/10.5555/3454287 (2019...

work page doi:10.18653/v1/2021.emnlp-main.236 2021

[10] [10]

& Bonini, N

Di Paola, C., Macheda, F., Laricchia, S., Weber, C. & Bonini, N. First-principles study of electronic transport and structural properties of Cu12Sb4S13 in its high-temperature phase. Phys. Rev. Res. 2, 033055 (2020). 59. Kresse, G. & Furthmüller, J. Efficient iterative schemes for ab initio total-energy calculations using a plane-wave basis set. Phys. Rev...

work page 2020

[11] [11]

Fan, Z. et al. Thermal conductivity decomposition in two-dimensional materials: Application to graphene. Phys. Rev. B 95, 144309 (2017). 70. Kubo, R. Statistical-Mechanical Theory of Irreversible Processes. I. General Theory and Simple Applications to Magnetic and Conduction Problems. https://doi.org/10.1143/JPSJ.12.570 12, 570–586 (2013). 71. Green, M. S...

work page doi:10.1143/jpsj.12.570 2017

[12] [12]

Saleemi, M. et al. Fabrication of nanostructured bulk Cobalt Antimonide (CoSb3) based skutterudites via bottom-up synthesis. MRS Online Proceedings Library 2013 1490:1 1490, 42–47 (2013). 80. Wei, M. et al. Enhanced Thermoelectric Performance of CoSb3 Thin Films by Ag and Ti Co-Doping. Materials 16, 1271 (2023). 81. Guo, R., Wang, X. & Huang, B. Thermal c...

work page 2013