Pith. sign in

REVIEW 3 major objections 5 minor 71 references

Predicting band gap from chemical composition: A simple learned model for a material property with atypical statistics

T0 review · 3 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A model with one learned parameter per chemical element predicts a crystal's electronic band gap from its composition alone, reaching a cross-validated mean absolute error of 0.575 eV by treating the gap as a mixed random variable with a…

desk verdict Simple ReLU composition model for band gaps is a useful heuristic, but the paper overclaims its statistical motivation without comparing the right baselines. read the letter →

arxiv 2501.02932 v1 pith:BZ4C7WTK submitted 2025-01-06 cond-mat.mtrl-sci cs.LGphysics.chem-ph

Andrew Ma , Owen Dugan , Marin Soljačić This is my paper · ORCID
classification cond-mat.mtrl-scics.LGphysics.chem-ph
keywords bandgappredictionchemicalcompositionmachinelearninginterpretablemodelmixedrandomvariableReLUelementparametersmaterialsinformatics
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that the electronic band gap of a crystalline solid can be predicted from its chemical formula alone, using a model simple enough to write on a periodic table. The authors first show that the gap is statistically atypical: its empirical distribution has a large discrete mass at exactly 0 eV (metals) plus a continuous positive part, so it is a mixed random variable rather than a standard continuous target. They therefore propose $\hat{\varepsilon}_{\rm relu}(M)=\mathrm{ReLU}\big(\sum_E w_E f_E(M)\big)$, one learned parameter $w_E$ per element, with a rectified linear unit clamping negative composition-weighted averages to zero. On 4,603 two-, three-, and four-element crystalline compounds, the model reaches a cross-validated test MAE of $0.575 \pm 0.036$ eV, beating the same model without the clamp ($0.824 \pm 0.035$ eV), and its learned element weights form a chemically sensible pattern. If this holds, a small lookup table gives instant, interpretable band gap estimates for materials whose crystal structure may not even be known.

What carries the argument

The load-bearing object is the ReLU composition model $\hat{\varepsilon}_{\rm relu}(M)=\max\big(0,\sum_E w_E f_E(M)\big)$, where $f_E(M)$ is the fraction of element $E$ in the chemical formula and $w_E$ is the single learned parameter for that element. The rectified linear unit (the maximum of the input and zero) is the mechanism that matches the target statistics: it enforces non-negativity and concentrates all negative pre-activations into a point mass at exactly 0 eV, mirroring the empirical distribution of the labels. The element-fraction representation is the only input — no crystal structure, bond geometry, or coordination enters — so the fitted model is entirely summarized by the vector of element weights.

What would settle it

Find one chemical formula with two known polymorphs whose measured or high-quality computed band gaps differ by more than about 0.575 eV (for example, carbon as diamond versus graphite); the model is forced to output the same prediction for both, so the larger of the two errors would exceed the claimed average and falsify the composition-only assumption.

Watch

Extended reading notes

Core claim

The paper's central claim is that band gap prediction should be treated as modeling a mixed random variable, and that the right simple model for this task is a composition-weighted average of per-element parameters followed by a ReLU: $\hat{\varepsilon}_{\rm relu}(M)=\mathrm{ReLU}\big(w\cdot f(M)\big)$. The discrete mass of metals at 0 eV is reproduced exactly by the clamp, and the continuous positive part is reproduced by the linear pre-activation for compounds whose weighted element score is positive. Trained by gradient descent on mean squared error over 4,603 materials and evaluated with 10-fold cross-validation, this model attains a test MAE of $0.575 \pm 0.036$ eV, outperforming the linear baseline $\hat{\varepsilon}_{\rm linear}(M)=w\cdot f(M)$ at $0.824 \pm 0.035$ eV. The learned weights, visualized on the periodic table, put positive values on the ionic, large-gap elements of the left and right columns and negative values on metallic transition elements, which the authors present as heuristic chemical interpretability: elements with larger parameters tend to form materials with larger band gaps.

Load-bearing premise

The load-bearing assumption is that a material's band gap is fully determined by its chemical formula through a fixed piecewise-linear rule, so that all structure-dependent physics such as bonding geometry, coordination, and polymorph identity can be ignored.

Editorial extensions

If this is right

  • A lookup table of element weights gives an instant band gap estimate for any 2-4 element crystalline solid, including materials with unknown crystal structure, at the reported cross-validated MAE of $0.575 \pm 0.036$ eV.
  • The ReLU clamp is doing real work: the same model without it has a test MAE of $0.824 \pm 0.035$ eV, makes negative predictions, and fails to reproduce the jump at 0 eV in the label distribution.
  • The learned element weights reproduce familiar chemistry — positive weights for ionic left- and right-side elements, negative weights for central transition metals — so the model doubles as a heuristic chemical rule for gap size.
  • Modeling the gap as a mixed random variable with a single model removes the need for the two-stage classify-then-regress pipeline sometimes used for band gap prediction.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct test of the chemical interpretation would be to regress the learned element weights against electronegativity values; a strong correlation would confirm the model has rediscovered a known heuristic rather than a dataset artifact.
  • Because the model outputs one gap per formula, its reported error averages over the training distribution of structures; for same-formula polymorphs with very different gaps, the error on at least one member of each pair is bounded below by half the gap difference, so the MAE should not be read as a per-material guarantee.
  • The same ReLU-of-weighted-average construction could be applied to other material properties with a physical floor and a large atom of observations at that floor; the accuracy gain from the clamp should scale with the size of the zero mass in the target distribution.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a composition-only model for the electronic band gap, epsilon_relu(M) = ReLU(sum_E w_E f_E(M)), with one learned weight per element. After documenting that the band gap distribution has a mass at 0 eV, the authors argue that a ReLU transform is suited to this mixed random variable. They fit the weights on a processed version of the Zhuo et al. dataset and report a 10-fold cross-validated MAE of 0.575 +/- 0.036 eV versus 0.824 +/- 0.035 eV for an ordinary least squares linear model. They also plot the empirical CDFs of predictions and present the learned weights on a periodic table as heuristic chemical interpretability.

Significance. If the central claim is established, the model is an attractive interpretable baseline and the mixed-random-variable framing is a useful corrective to treating the band gap as an ordinary continuous target. The paper's strengths are its simplicity, the use of cross-validation, the explicit empirical-CDF analysis, and the fact that the model has only one parameter per element. However, the specific claim that the ReLU form is responsible for capturing the atypical statistics is not yet supported, because the paper does not test simpler non-negativity baselines or compare with existing composition-only models. The reported MAE values are internally consistent with the described training procedure, but the load-bearing comparison is incomplete.

major comments (3)
  1. [Section IV, Table I] The reported difference between Eq. (8) and Eq. (9) does not establish that the ReLU nonlinearity, rather than mere non-negativity, drives the improvement. Since 53.2% of the labels are exactly 0 eV and the ReLU clamps every negative pre-activation to zero, the MAE gain can be reproduced in large part by the positive part of the OLS predictor, max(0, w_OLS dot f(M)); this baseline is not reported. The authors should also report a zero-inflated baseline (e.g., Tweedie regression or a two-stage classifier plus regressor). If the positive-part baseline achieves MAE close to 0.575 eV, then the claim that the ReLU form is specifically suited to the band gap's mixed statistics is unsupported.
  2. [Section IV, Fig. 3] The jump in the ReLU model's prediction eCDF at 0 eV is partly by construction: for any composition with w dot f(M) <= 0, Eq. (8) returns exactly zero, so the model is guaranteed to produce a point mass at zero whenever a non-negligible fraction of the test set falls in that halfspace. Using this jump as evidence that the model captures the mixed nature of the target is therefore circular unless compared against a non-negativity-enforcing baseline. The manuscript should quantify the similarity of predicted and label eCDFs (e.g., Kolmogorov-Smirnov distance) for the ReLU model and for max(0, linear), and report the fraction of test predictions exactly equal to zero for each.
  3. [Section I and Section IV] The paper does not compare its claimed MAE with published composition-only band gap models, despite citing several such works in Section I (e.g., Zhuo et al. 2018 [39], Venkatraman 2021 [38], Goodall and Lee 2020 [40], CrabNet [42,44]). Without at least one quantitative comparison on the same data and splits, the practical significance of the 0.575 eV MAE is unclear; the model could be either a strong simple baseline or far worse than existing composition-based methods. A table with published MAE values on the same (or closely matched) dataset would make the claim concrete.
minor comments (5)
  1. [Section II] There is a typo in the sentence 'the band gap is actually not a continous random variable'; 'continous' should be 'continuous'.
  2. [Section III, last paragraph] The dataset is described only in Supplementary Section S1; the main text should also state the number of materials (4,603), the 53.2% fraction of zero-gap labels, the restriction to 2-4 element materials, and the provenance of labels (experimental versus Materials Project DFT) so that the reported MAE is interpretable without consulting the supplement.
  3. [Section IV, Fig. 3] The green shaded region is not defined in the caption beyond 'standard deviation'; please specify whether it is the standard deviation of the eCDF across cross-validation folds at each point or a pointwise interval of another kind.
  4. [Code Availability] The Code Availability section says the code 'will be made available in a public repository' but provides no repository link or DOI; for a reproducible machine learning study, a URL or accession code should be provided at submission time.
  5. [Abstract and Section III] The abstract and title claim 'predicting band gap from chemical composition' without mentioning the 2-4 element restriction and the exclusion of pure-element materials; a brief caveat would prevent overgeneralization of the model's scope.

Circularity Check

2 steps flagged · score 4.0 of 10

The ReLU model's zero-mass jump and its MAE advantage over the linear baseline are largely built into Eq. (8); the cross-validated absolute MAE remains an independent result.

  1. self definitional [Section III, Eq. (8); Section IV, Fig. 3]
    "The intention behind this ReLU formulation is to design a model that is tailored for the non-standard statistics of the band gap. ... the ReLU model captures the existence of a discrete probability mass at 0 eV, whereas the baseline linear model does not."

    By Eq. (8), epsilon_relu(M) = ReLU(w·f(M)) is exactly 0 for every composition in the halfspace {M : w·f(M) <= 0}. A point mass at 0 is therefore a logical consequence of the chosen nonlinearity, present for any fitted weights, not an empirical discovery. The paper uses this guaranteed property as evidence that the model achieves its design goal of matching the band gap's mixed-distribution statistics; only the height of the mass is data-dependent, while its existence is definitional.

  2. other [Section IV, Table I]
    "Comparing the test MAE of the ReLU model with the test MAE of the baseline linear model illustrates that the use of the ReLU transform provides significant benefits in terms of reducing error."

    For any nonnegative band gap label epsilon and any real prediction x, |epsilon - max(0,x)| <= |epsilon - x|, because clamping x to 0 moves the prediction toward every nonnegative target. Hence the ReLU model class (Eq. 8) dominates the linear model class (Eq. 9) pointwise in absolute error before any fitting is done. The reported advantage is therefore a mathematically guaranteed consequence of enforcing non-negativity rather than specific evidence for the 'atypical statistics' / mixed-random-variable design; a proper test would need zero-inflated baselines such as Tweedie or two-stage models.

full rationale

The paper's central quantitative result, a 0.575 +/- 0.036 eV cross-validated MAE from element fractions, is genuine held-out prediction and is not circular: parameters are fitted by MSE on training folds and evaluated on held-out materials. The interpretability claims are explicitly heuristic. The citations to the authors' earlier classification work [46,47] motivate the one-parameter-per-element ansatz but are not load-bearing, because the model is fully specified in Eqs. (8)-(9) and its parameters are refit here. Two local pieces of evidence are, however, partly tautological. First, the ReLU model's point mass at 0 eV is guaranteed by Eq. (8) (any composition with w·f <= 0 maps to 0), so Fig. 3's 'capture' of the zero-mass feature is by construction; only the mass height is informative. Second, because clamping any number toward 0 can only reduce absolute error against a nonnegative target, the ReLU MAE advantage over the unclamped linear baseline in Table I is mathematically guaranteed; it does not uniquely confirm the 'mixed random variable' design without additional zero-inflated baselines. These points weaken a supporting comparison, but they do not make the primary band-gap predictions circular.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claim rests on the fitted per-element weights and on modeling assumptions about the sufficiency of composition and the adequacy of a fixed zero cutoff. No new physical entities are introduced.

free parameters (1)
  • Element weight w_E for each element = e.g., w_Be = 2.07 eV, w_Ti = -8.19 eV, w_Sb = -1.22 eV, w_Se = 3.07 eV, w_Br = 3.57 eV (Fig. 4)
    Each element in the dataset receives a scalar parameter fit by minimizing MSE on the Zhuo et al. data. The model prediction is a weighted average of these fitted values, so the central result depends directly on them.
assumptions (3)
  • domain assumption Band gap labels in the training set are accurate enough for supervised learning, mixing experimental values for nonmetals and DFT values for metals.
    The paper uses the processed Zhuo et al. dataset without quantifying label noise; mixed experimental and DFT sources may introduce systematic errors. This enters in Section II and Supplementary S1.
  • domain assumption Chemical composition alone, through element fractions, carries enough information to predict the band gap; crystal structure is ignored.
    The model input is only the formula in Eq. (8); all structural information is discarded, so structure-dependent gaps such as allotropes or polymorphs cannot be represented.
  • ad hoc to paper The ReLU form with a fixed threshold at zero is an adequate model for the zero and nonzero gap mixture.
    The threshold is fixed at zero and is chosen to reproduce the measured probability mass at 0 eV; the paper does not compare against alternative mixed-distribution models such as two-stage classifiers.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Predicting band gap from chemical composition: A simple learned model for a material property with atypical statistics." pith.science (2026). https://pith.science/paper/BZ4C7WTK

@misc{pith2026250102932,
  author       = {Pith},
  title        = {Pith review of: Predicting band gap from chemical composition: A simple learned model for a material property with atypical statistics},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BZ4C7WTK}},
  note         = {Machine review of arXiv:2501.02932}
}
read the original abstract

In solid-state materials science, substantial efforts have been devoted to the calculation and modeling of the electronic band gap. While a wide range of ab initio methods and machine learning algorithms have been created that can predict this quantity, the development of new computational approaches for studying the band gap remains an active area of research. Here we introduce a simple machine learning model for predicting the band gap using only the chemical composition of the crystalline material. To motivate the form of the model, we first analyze the empirical distribution of the band gap, which sheds new light on its atypical statistics. Specifically, our analysis enables us to frame band gap prediction as a task of modeling a mixed random variable, and we design our model accordingly. Our model formulation incorporates thematic ideas from chemical heuristic models for other material properties in a manner that is suited towards the band gap modeling task. The model has exactly one parameter corresponding to each element, which is fit using data. To predict the band gap for a given material, the model computes a weighted average of the parameters associated with its constituent elements and then takes the maximum of this quantity and zero. The model provides heuristic chemical interpretability by intuitively capturing the associations between the band gap and individual chemical elements.

Figures

Figures reproduced from arXiv: 2501.02932 by the authors.

Figure 1
Figure 1. The atypical distribution of the band gap. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Modeling approach based on one parameter for each chemical element. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Empirical distribution of model predictions [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Periodic table visualization of the learned parameters [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

71 extracted references · 70 canonical work pages

  1. [39]

    Y . Zhuo, A. Mansouri Tehrani, and J. Brgoch, Predicting the band gaps of inorganic solids by machine learning, The Journal of Physical Chemistry Letters 9, 1668 (2018)

    work page 2018

  2. [38]

    Venkatraman, The utility of composition-based machine learning models for band gap prediction, Computational Materi- als Science 197, 110637 (2021)

    V . Venkatraman, The utility of composition-based machine learning models for band gap prediction, Computational Materi- als Science 197, 110637 (2021)

    work page 2021

  3. [40]

    R. E. Goodall and A. A. Lee, Predicting materials properties without crystal structure: Deep representation learning from stoichiometry, Nature Communications 11, 1 (2020)

    work page 2020

  4. [1]

    J. P. Perdew, Density functional theory and the band gap prob- lem, International Journal of Quantum Chemistry 28, 497 (1985)

    work page 1985

  5. [2]

    Sham and M

    L. Sham and M. Schl¨uter, Density-functional theory of the band gap, Physical Review B 32, 3883 (1985)

    work page 1985

  6. [3]

    Tran and P

    F. Tran and P. Blaha, Accurate band gaps of semiconductors and insulators with a semilocal exchange-correlation potential, Physical Review Letters 102, 226401 (2009)

    work page 2009

  7. [4]

    M. K. Chan and G. Ceder, E fficient band gap prediction for solids, Physical Review Letters 105, 196403 (2010)

    work page 2010

  8. [5]

    Bagayoko, Understanding density functional theory (DFT) and completing it in practice, AIP Advances 4 (2014)

    D. Bagayoko, Understanding density functional theory (DFT) and completing it in practice, AIP Advances 4 (2014)

    work page 2014

Show all 71 references
  1. [6]

    H. Xiao, J. Tahir-Kheli, and W. A. Goddard III, Accurate band gaps for semiconductors from density functional theory, The Journal of Physical Chemistry Letters 2, 212 (2011)

    2011

  2. [7]

    A. J. Garza and G. E. Scuseria, Predicting band gaps with hybrid density functionals, The Journal of Physical Chemistry Letters 7, 4165 (2016)

    2016

  3. [8]

    Borlido, T

    P. Borlido, T. Aull, A. W. Huran, F. Tran, M. A. Marques, and S. Botti, Large-scale benchmark of exchange–correlation func- tionals for the determination of electronic band gaps of solids, Journal of Chemical Theory and Computation 15, 5069 (2019)

    2019

  4. [9]

    S. Kim, M. Lee, C. Hong, Y . Yoon, H. An, D. Lee, W. Jeong, D. Yoo, Y . Kang, Y . Youn,et al. , A band-gap database for semiconducting inorganic materials calculated with hybrid func- tional, Scientific Data 7, 387 (2020)

    2020

  5. [10]

    A. Jain, S. P. Ong, G. Hautier, W. Chen, W. D. Richards, S. Dacek, S. Cholia, D. Gunter, D. Skinner, G. Ceder, et al., Commentary: The Materials Project: A materials genome ap- proach to accelerating materials innovation, APL materials 1 (2013)

    2013

  6. [11]

    Choudhary, K

    K. Choudhary, K. F. Garrity, A. C. Reid, B. DeCost, A. J. Bi- acchi, A. R. Hight Walker, Z. Trautt, J. Hattrick-Simpers, A. G. Kusne, A. Centrone, et al., The joint automated repository for various integrated simulations (JARVIS) for data-driven materi- als design, npj Computa...

    2020

  7. [12]

    Curtarolo, W

    S. Curtarolo, W. Setyawan, G. L. Hart, M. Jahnatek, R. V . Chep- ulskii, R. H. Taylor, S. Wang, J. Xue, K. Yang, O. Levy,et al., AFLOW: An automatic framework for high-throughput materi- als discovery, Computational Materials Science 58, 218 (2012)

    2012

  8. [13]

    D. Wing, G. Ohad, J. B. Haber, M. R. Filip, S. E. Gant, J. B. Neaton, and L. Kronik, Band gaps of crystalline solids from wannier-localization–based optimal tuning of a screened range-separated hybrid functional, Proceedings of the National Academy of Sciences 118, e2104556118 (2021)

    2021

  9. [14]

    M. Liu, A. Gopakumar, V . I. Hegde, J. He, and C. Wolver- ton, High-throughput hybrid-functional DFT calculations of bandgaps and formation energies and multifidelity learning with uncertainty quantification, Physical Review Materials8, 043803 (2024)

    2024

  10. [15]

    K. T. Butler, D. W. Davies, H. Cartwright, O. Isayev, and A. Walsh, Machine learning for molecular and materials science, Nature 559, 547 (2018)

    2018

  11. [16]

    Himanen, A

    L. Himanen, A. Geurts, A. S. Foster, and P. Rinke, Data-driven materials science: status, challenges, and perspectives, Ad- 8 vanced Science 6, 1900808 (2019)

    2019

  12. [17]

    G. R. Schleder, A. C. Padilha, C. M. Acosta, M. Costa, and A. Fazzio, From DFT to machine learning: recent approaches to materials science—a review, Journal of Physics: Materials 2, 032001 (2019)

    2019

  13. [18]

    Damewood, J

    J. Damewood, J. Karaguesian, J. R. Lunger, A. R. Tan, M. Xie, J. Peng, and R. G´omez-Bombarelli, Representations of materials for machine learning, Annual Review of Materials Research 53, 399 (2023)

    2023

  14. [19]

    Xie and J

    T. Xie and J. C. Grossman, Crystal graph convolutional neural networks for an accurate and interpretable prediction of material properties, Physical Review Letters 120, 145301 (2018)

    2018

  15. [20]

    K. T. Sch¨utt, H. E. Sauceda, P.-J. Kindermans, A. Tkatchenko, and K.-R. M ¨uller, Schnet–a deep learning architecture for molecules and materials, The Journal of Chemical Physics 148 (2018)

    2018

  16. [21]

    C. Chen, W. Ye, Y . Zuo, C. Zheng, and S. P. Ong, Graph net- works as a universal machine learning framework for molecules and crystals, Chemistry of Materials 31, 3564 (2019)

    2019

  17. [22]

    Choudhary and B

    K. Choudhary and B. DeCost, Atomistic line graph neural net- work for improved materials property predictions, npj Compu- tational Materials 7, 185 (2021)

    2021

  18. [23]

    Y . Lin, K. Yan, Y . Luo, Y . Liu, X. Qian, and S. Ji, Efficient approximations of complete interatomic potentials for crystal property prediction, in Proceedings of the 40th International Conference on Machine Learning , Proceedings of Machine Learning Research, V ol. 202, ed...

  19. [24]

    V . Moro, C. Loh, R. Dangovski, A. Ghorashi, A. Ma, Z. Chen, S. Kim, P. Y . Lu, T. Christensen, and M. Soljaˇci´c, Multimodal learning for materials, arXiv preprint arXiv:2312.00111v3 (2024)

    2024 arXiv

  20. [25]

    X.-G. Li, B. Blaiszik, M. E. Schwarting, R. Jacobs, A. Scourtas, K. Schmidt, P. M. V oyles, and D. Morgan, Graph network based deep learning of bandgaps, The Journal of Chemical Physics 155 (2021)

    2021

  21. [26]

    H. Tu, Y . Han, Z. Wang, A. Chen, K. Tao, S. Ye, S. Wang, Z. Wei, and J. Li, Rotnet: A rotationally invariant graph neural network for quantum mechanical calculations, Small Methods 8, 2300534 (2024)

    2024

  22. [27]

    Batzner, A

    S. Batzner, A. Musaelian, L. Sun, M. Geiger, J. P. Mailoa, M. Kornbluth, N. Molinari, T. E. Smidt, and B. Kozinsky, E(3)- equivariant graph neural networks for data-efficient and accu- rate interatomic potentials, Nature Communications 13, 2453 (2022)

    2022

  23. [28]

    Xie and J

    T. Xie and J. C. Grossman, Hierarchical visualization of materi- als space with graph convolutional neural networks, The Journal of Chemical Physics 149 (2018)

    2018

  24. [29]

    Z. Chen, N. Andrejevic, T. Smidt, Z. Ding, Q. Xu, Y .-T. Chi, Q. T. Nguyen, A. Alatas, J. Kong, and M. Li, Direct prediction of phonon density of states with euclidean neural networks, Advanced Science 8, 2004214 (2021)

    2021

  25. [30]

    C. W. Park and C. Wolverton, Developing an improved crys- tal graph convolutional neural network framework for acceler- ated materials discovery, Physical Review Materials4, 063801 (2020)

    2020

  26. [31]

    V . Fung, P. Ganesh, and B. G. Sumpter, Physically informed ma- chine learning prediction of electronic density of states, Chem- istry of Materials 34, 4848 (2022)

    2022

  27. [32]

    P. Manca, A relation between the binding energy and the band- gap energy in semiconductors of diamond or zinc-blende struc- ture, Journal of Physics and Chemistry of Solids 20, 268 (1961)

    1961

  28. [33]

    Xiuying, G

    G. Xiuying, G. Fengsheng, T. Yamaguchi, H. Kan, and M. Ku- magawa, Dependence of energy band gap and lattice constant of III-V semiconductors on electronegativity difference of the con- stituent elements, Crystal Research and Technology 27, 1087 (1992)

    1992

  29. [34]

    A. V . Prokofiev, A. I. Shelykh, and B. T. Melekh, Periodicity in the band gap variation ofLn2X3 (X = O, S, Se) in the lanthanide series, Journal of Alloys and Compounds 242, 41 (1996)

    1996

  30. [35]

    K. L. Heng, S. J. Chua, and P. Wu, Prediction of semiconduc- tor material properties by the properties of their constituent chemical elements, Chemistry of Materials 12, 1648 (2000)

    2000

  31. [36]

    K. Li, Y . Li, and D. Xue, Band gap engineering of crystal materi- als: band gap estimation of semiconductors via electronegativity, Functional Materials Letters 5, 1260002 (2012)

    2012

  32. [37]

    K ¨orbel, M

    S. K ¨orbel, M. A. L. Marques, and S. Botti, Stability and electronic properties of new inorganic perovskites from high- throughput ab initio calculations, Journal of Materials Chemistry C 4, 3157 (2016)

    2016

  33. [41]

    L. Ward, A. Agrawal, A. Choudhary, and C. Wolverton, A general-purpose machine learning framework for predicting properties of inorganic materials, npj Computational Materials 2, 1 (2016)

    2016

  34. [42]

    A. Y .-T. Wang, S. K. Kauwe, R. J. Murdock, and T. D. Sparks, Compositionally restricted attention-based network for mate- rials property predictions, npj Computational Materials 7, 77 (2021)

    2021

  35. [43]

    Prein, E

    T. Prein, E. Pan, T. Doerr, E. Olivetti, and J. L. Rupp, MTEN- CODER: A multi-task pretrained transformer encoder for ma- terials representation learning, in AI for Accelerated Materials Design-NeurIPS 2023 Workshop (2023)

    2023

  36. [44]

    A. Y .-T. Wang, M. S. Mahmoud, M. Czasny, and A. Gurlo, CrabNet for explainable deep learning in materials science: bridging the gap between academia and industry, Integrating Materials and Manufacturing Innovation 11, 41 (2022)

    2022

  37. [45]

    Huang, R

    H. Huang, R. Magar, and A. Barati Farimani, Pretraining strate- gies for structure agnostic material property prediction, Journal of Chemical Information and Modeling 64, 627 (2024)

    2024

  38. [46]

    A. Ma, Y . Zhang, T. Christensen, H. C. Po, L. Jing, L. Fu, and M. Soljaˇci´c, Topogivity: A machine-learned chemical rule for discovering topological materials, Nano Letters 23, 772 (2023)

    2023

  39. [47]

    Ma and M

    A. Ma and M. Soljaˇci´c, Learning chemical heuristics for classi- fying solid-state materials: Parameter interpretability and induc- tive biases, In Preparation (2025)

    2025

  40. [48]

    A. Dunn, Q. Wang, A. Ganose, D. Dopp, and A. Jain, Bench- marking materials property prediction methods: The Matbench test set and Automatminer reference algorithm, npj Computa- tional Materials 6, 138 (2020)

    2020

  41. [49]

    Hastie, R

    T. Hastie, R. Tibshirani, and J. Friedman, The Elements of Statistical Learning: Data Mining, Inference, and Prediction (Springer Science & Business Media, 2009)

    2009

  42. [50]

    A. W. Van der Vaart,Asymptotic Statistics, V ol. 3 (Cambridge University Press, 2000)

    2000

  43. [51]

    Coles, J

    S. Coles, J. Bawa, L. Trenner, and P. Dorazio,An Introduction to Statistical Modeling of Extreme V alues, V ol. 208 (Springer, 2001)

    2001

  44. [52]

    Belitz and P

    K. Belitz and P. Stackelberg, Evaluation of six methods for cor- recting bias in estimates from ensemble tree machine learning regression models, Environmental Modelling & Software 139, 9 105006 (2021)

    2021

  45. [53]

    Kingma and J

    D. Kingma and J. Ba, Adam: A method for stochastic optimiza- tion, in International Conference on Learning Representations (2015)

    2015

  46. [54]

    C. Yang, X. Chong, M. Hu, W. Yu, J. He, Y . Zhang, J. Feng, Y . Zhou, and L.-W. Wang, Accelerating the discovery of hybrid perovskites with targeted band gaps via interpretable machine learning, ACS Applied Materials & Interfaces 15, 40419 (2023)

    2023

  47. [55]

    Isayev, C

    O. Isayev, C. Oses, C. Toher, E. Gossett, S. Curtarolo, and A. Tropsha, Universal fragment descriptors for predicting prop- erties of inorganic crystals, Nature Communications 8, 15679 (2017)

    2017

  48. [56]

    Michaelis, N

    P. Michaelis, N. Klein, and T. Kneib, Mixed discrete-continuous regression—a novel approach based on weight functions, Stat 9, e277 (2020)

    2020

  49. [57]

    Lambert, Zero-inflated poisson regression, with an applica- tion to defects in manufacturing, Technometrics 34, 1 (1992)

    D. Lambert, Zero-inflated poisson regression, with an applica- tion to defects in manufacturing, Technometrics 34, 1 (1992)

    1992

  50. [58]

    Jørgensen and M

    B. Jørgensen and M. C. Paes De Souza, Fitting tweedie’s com- pound poisson model to insurance claims data, Scandinavian Actuarial Journal 1994, 69 (1994)

    1994

  51. [59]

    H. Zhou, W. Qian, and Y . Yang, Tweedie gradient boosting for extremely unbalanced zero-inflated data, Communications in Statistics-Simulation and Computation 51, 5507 (2022)

    2022

  52. [60]

    Choubisa, P

    H. Choubisa, P. Todorovi ´c, J. M. Pina, D. H. Parmar, Z. Li, O. V oznyy, I. Tamblyn, and E. H. Sargent, Interpretable dis- covery of semiconductors with machine learning, npj Computa- tional Materials 9, 117 (2023)

    2023

  53. [61]

    C. J. Bartel, C. Sutton, B. R. Goldsmith, R. Ouyang, C. B. Musgrave, L. M. Ghiringhelli, and M. Scheffler, New tolerance factor to predict the stability of perovskite oxides and halides, Science Advances 5, eaav0693 (2019)

    2019

  54. [62]

    Y . Liu, M. Jovanovic, K. Mallayya, W. J. Maddox, A. G. Wilson, S. Klemenz, L. M. Schoop, and E.-A. Kim, Materials expert- artificial intelligence for materials discovery, arXiv preprint arXiv:2312.02796 (2023)

    2023

  55. [63]

    Ouyang, S

    R. Ouyang, S. Curtarolo, E. Ahmetcik, M. Scheffler, and L. M. Ghiringhelli, Sisso: A compressed-sensing method for identi- fying the best low-dimensional descriptor in an immensity of offered candidates, Physical Review Materials2, 083802 (2018)

    2018

  56. [64]

    A. E. Allen and A. Tkatchenko, Machine learning of material properties: Predictive and interpretable multilinear models, Sci- ence Advances 8, eabm7185 (2022)

    2022

  57. [65]

    E. S. Muckley, J. E. Saal, B. Meredig, C. S. Roper, and J. H. Mar- tin, Interpretable models for extrapolation in scientific machine learning, Digital Discovery 2, 1425 (2023)

    2023

  58. [66]

    X. Chen, S. Lu, X. Wan, Q. Chen, Q. Zhou, and J. Wang, Ac- curate property prediction with interpretable machine learning model for small datasets via transformed atom vector, Physical Review Materials 6, 123803 (2022)

    2022

  59. [67]

    Zhang, Y .-M

    X. Zhang, Y .-M. Zhao, Z. Song, and L. Shen, Physically explain- able statistical learning of flat bands in stoichiometric materials from the periodic table, Physical Review Materials 7, 064804 (2023)

    2023

  60. [68]

    Ziletti, D

    A. Ziletti, D. Kumar, M. Scheffler, and L. M. Ghiringhelli, In- sightful classification of crystal structures using deep learning, Nature Communications 9, 2775 (2018)

    2018

  61. [69]

    George and G

    J. George and G. Hautier, Chemist versus machine: Traditional knowledge versus machine learning techniques, Trends Chem. 3, 86 (2020)

    2020

  62. [70]

    Oviedo, J

    F. Oviedo, J. L. Ferres, T. Buonassisi, and K. T. Butler, Inter- pretable and explainable machine learning for materials science and chemistry, Accounts of Materials Research 3, 597 (2022)

    2022

  63. [71]

    S. Huo, S. Zhang, Q. Wu, and X. Zhang, Feature-assisted ma- chine learning for predicting band gaps of binary semiconduc- tors, Nanomaterials 14, 445 (2024)

    2024

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.