Learning disentangled latent representations facilitates discovery and design of functional materials

Jaehoon Cha; Keith T. Butler; Matthew Walker; Tingyao Lu

REVIEW 4 major objections 5 minor 43 references

Learning disentangled latent representations facilitates discovery and design of functional materials

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read Unsupervised autoencoder finds solar materials without labels

desk verdict Solid unsupervised demo with a nice SLME-correlated latent dimension, but the discovery claim needs better robustness checks and the abstract overstates reconstruction fidelity. read the letter →

arxiv 2507.19602 v1 pith:4W2UOCHW submitted 2025-07-25 cond-mat.mtrl-sci

Jaehoon Cha , Tingyao Lu , Matthew Walker , Keith T. Butler This is my paper

classification cond-mat.mtrl-sci

keywords disentangledautoencodermaterialsdiscoveryphotovoltaicSLMEopticalabsorptionspectraunsupervisedrepresentationlearningdirect-to-indirectbandgaplatentspacesearch

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 claims that a disentangling autoencoder trained only on raw optical absorption spectra, with no efficiency labels, learns a latent dimension that tracks the Spectroscopic Limited Maximum Efficiency (SLME) of photovoltaic materials, reaching a Pearson correlation of -0.76. That dimension encodes the shape of the absorption onset and corresponds to the transition from direct to indirect optical band gaps, a known determinant of absorber quality. Because the representation is compact and interpretable, a nearest-neighbour search in the latent space starting from a known high-SLME material recovers the top-20 SLME materials after exploring about 43% of the candidate pool, outperforming a beta-VAE and random sampling. The practical stake is that unsupervised structure learning could reduce the need for expensive labelled screening in materials discovery.

What carries the argument

The load-bearing object is the Disentangling Autoencoder, an autoencoder architecture that enforces orthogonality in the latent space through normalisation, interpolation, and a layer that constrains decoder output variations to be orthogonal across latent dimensions. The paper adapts this architecture to one-dimensional optical spectra, using a nine-dimensional latent space and training only with a reconstruction loss (mean absolute error). The argument runs through two mechanisms working together: the orthogonality constraint isolates an interpretable spectral feature (direct versus indirect band gap), and a nearest-neighbour search on the resulting latent space supplies the discovery ordering. The search ranks all candidate materials by Manhattan distance from the single known top-SLME material in the dataset.

What would settle it

Train the DAE on one half of the spectra and run the latent-space discovery search on the held-out half; if the search does not recover top-SLME materials faster than random sampling on that held-out set, the claimed discovery advantage is an artifact of evaluating the search on the training data.

Watch

Extended reading notes

Core claim

The central discovery is that a nine-dimensional latent space, learned from 17,283 simulated optical absorption spectra using only a reconstruction loss, contains a single dimension whose variation is strongly and negatively correlated with SLME (r = -0.76), even though SLME labels were never used in training. Traversing this dimension shifts the absorption onset from concave (high SLME, direct band gap) to convex (low SLME, indirect band gap). The same dimension stands out in a Lasso regression as the only coefficient needed to predict SLME, whereas a beta-VAE and PCA spread predictive weight across many entangled features. In a simulated discovery campaign, ranking candidates by Manhattan distance from the top-SLME material in the DAE latent space finds all 20 top performers within about 43% of the search space, which the authors interpret as more efficient than VAE-guided and random search.

Load-bearing premise

The discovery campaign assumes that Manhattan distance in the DAE latent space from the single highest-SLME material orders the remaining materials by photovoltaic potential, and the search is evaluated on the same data used to train the representation.

Editorial extensions

If this is right

Because the DAE recovers the direct-to-indirect band gap signal without labels, SLME labels may not be needed to prioritise candidate absorber materials from raw spectra.
The DAE latent space finds all top-20 SLME materials after exploring roughly 43% of the candidate pool, whereas random sampling is far less efficient.
A Lasso model predicts SLME from the DAE latent variables using essentially one dimension, indicating a compact and interpretable encoding of the efficiency-relevant physics.
The DAE reconstructs fine spectral features, such as the absorption bump in PbSe, more accurately than the beta-VAE, which smooths them out.
The same unsupervised approach could be applied to other high-dimensional materials signals, such as X-ray diffraction patterns or density of states, where labels are scarce.

Reading between the lines

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

A stronger test of the representation would freeze the trained encoder and measure the latent dimension's correlation with SLME on a held-out set of spectra, separating genuine representation quality from memorization of the training distribution.
The discovery result may depend on the chosen seed material and the Manhattan distance metric; testing multiple seeds and alternative metrics would show whether the DAE advantage is robust or specific to this single setup.
Because the latent space is only nine-dimensional and ordered by distance to a known good absorber, it could serve as a cheap pre-filter in a high-throughput pipeline before expensive DFT or SLME calculations.
Traversing the SLME-correlated dimension could act as a generative design axis, proposing spectra with engineered direct-gap character, which the paper suggests but does not demonstrate.

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.

Desk Editor's Note

read the letter

Short version: the DAE application to 17k optical absorption spectra is a real, reproducible piece of work with open code and data, and the finding that one latent dimension tracks SLME (r = -0.76) and corresponds to the direct/indirect gap transition is genuinely interesting. But the paper's headline discovery claim is built on a single, under-tested protocol, and the abstract contains a direct contradiction with the paper's own numbers.

What's actually new: applying the authors' DAE architecture to 1D spectra, the single SLME-correlated latent dimension, and the simulated discovery campaign. The Lasso analysis adds support: DAE uses one dimension for SLME prediction, while VAE/PCA spread it out. The latent traversal figures are convincing. This is not just a re-run; the physical interpretation is plausible and opens a useful direction.

Where it's soft: first, the abstract and conclusion say DAE achieves 'superior reconstruction fidelity,' but the MSE numbers in the text show PCA at 0.00004 vs DAE at 0.00010. The text admits PCA is lower but calls the difference negligible; the abstract doesn't. That's an overstatement that should be fixed. Second, the discovery campaign (Figure 2) is one protocol: Manhattan distance from a single top-SLME seed, on the full training set, with no error bars on the DAE/VAE/PCA curves and no plain autoencoder baseline. The stress-test is right that the ranking could be sensitive to metric or seed. Also, the text credits PCA in the discovery results but the Figure 2 caption doesn't list PCA—another inconsistency. Third, the direct/indirect gap interpretation is inferred from spectral shape, not validated against known band-structure labels; that's a minor point, but it would help to check a few materials.

None of this sinks the core unsupervised result. The SLME correlation is measured post hoc, so there's no circularity, and the code and data are available for verification. The paper is worth refereeing, but the authors should fix the abstract, add robustness checks (different distances, seeds, a held-out split, error bars), and reconcile the PCA mention with the figure. If you're screening for interesting representation-learning results in materials, read it; just don't use the discovery curves as evidence until they're hardened.

Referee Report

4 major / 5 minor

Summary. The paper trains a Disentangling Autoencoder (DAE) on 17,283 simulated optical absorption spectra without using SLME labels, and compares it with a beta-VAE and PCA in a nine-dimensional latent space. The authors report that one DAE latent dimension correlates strongly with SLME (Pearson r = -0.76), that this dimension corresponds to a direct-to-indirect band-gap transition, that DAE representations are more compact and predictive for SLME than the baselines, and that a latent-space nearest-neighbor search seeded by the top-SLME material discovers the top-20 SLME materials more efficiently than random search or VAE-guided search. The central scientific claim is that unsupervised, disentangled representations of raw spectra can recover physically meaningful, efficiency-relevant structure and accelerate materials discovery.

Significance. If the claims hold, this would be a valuable demonstration of unsupervised representation learning for materials discovery, with a physically interpretable latent dimension and an open implementation. Strengths of the paper include the post-hoc evaluation of the SLME correlation rather than using labels during training, the use of a held-out test split for the Lasso prediction task, and the availability of code and data for reproducibility. However, the paper currently overstates reconstruction fidelity relative to its own numbers, and the discovery-advantage claim rests on a single evaluation protocol with limited robustness checks. The physical interpretation, while plausible, is not independently validated against band-structure labels.

major comments (4)

[Abstract and Section III.C] The abstract states that the DAE achieves 'superior reconstruction fidelity' relative to PCA and beta-VAE, and the conclusion repeats that the DAE 'not only achieved lower reconstruction error.' These statements are contradicted by the MSE values reported in Section III.C: DAE 0.00010, beta-VAE 0.00125, and PCA 0.00004. PCA has the lowest reconstruction error, so the claim should be corrected or explicitly qualified as applying only to nonlinear methods or to reconstruction quality at visually comparable fidelity.
[Section III.A and Figure 2] The discovery-advantage claim rests on a single protocol: all materials are ranked by Manhattan distance from the single top-SLME material in the nine-dimensional latent space, and the search is evaluated on the same full dataset used to train the representation. Figure 2 shows a +/-1 sigma band only for random search; the DAE, VAE, and PCA curves have no uncertainty or seed variation, and the figure caption does not list PCA even though the text reports PCA results in the same section. Please add error bars or seed variation, test alternative distance metrics and alternative seed materials, and evaluate on a held-out split. Without these robustness checks, the headline 'more efficient discovery' claim is not yet secure.
[Section III.B and Figure 5] The physical interpretation that the SLME-correlated latent dimension encodes the direct-to-indirect band-gap transition is inferred solely from the concave-to-convex evolution of the absorption onset during latent traversal. Since SLME is itself computed from these same absorption spectra (Section II), a correlation between an onset-shape feature and SLME is partly expected. To support the claim that the DAE has discovered a physically meaningful disentangled factor, the authors should validate the interpretation against independent direct/indirect band-structure labels for the traversed materials, or at least compare the latent feature against a simple handcrafted onset-shape descriptor as a baseline.
[Section V (Methods)] The comparison does not include a plain autoencoder baseline with the same architecture but without the orthogonality-enforcing components, nor does it report sensitivity to the chosen latent dimensionality (nine) and hyperparameters (alpha = 0.005, beta = 2). Because the DAE's claimed advantages are attributed to its architectural disentanglement mechanism, the absence of such controls leaves open the possibility that a standard autoencoder with the same capacity would perform equally well on reconstruction and discovery.

minor comments (5)

[Section III.A and Figure 2 caption] The text in Section III.A states that 'DAE and PCA have recovered more than 60% of the top 20 materials,' but Figure 2's caption lists only DAE, VAE, and random; the PCA curve should either be added to the figure or removed from the text.
[Figure 3 caption] There is a typo in the caption: 'coeffieicnts' should be 'coefficients.'
[Section II and Section III.A] The manuscript gives both 17,283 materials (Section II) and 17,282 candidates (Section III.A, '7500 out of 17282'); please reconcile these numbers.
[Section III.C] The MSE values are reported without units or normalization context; since the absorption values are normalized, this should be stated explicitly, and it would be helpful to report whether these are computed on the same train/test split as the Lasso evaluation.
[Section V (Methods)] The text says the DAE and beta-VAE share the same architecture, but then states that the DAE uses a single fully connected layer while the beta-VAE uses two fully connected layers; please clarify whether this architectural difference is intentional and how it affects the comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the DAE is trained unsupervised on spectra, and the SLME correlation is measured post hoc rather than fitted.

full rationale

The paper's central derivation is self-contained. The DAE is trained solely on optical absorption spectra using a reconstruction loss, with no SLME labels in the training objective (Section V, Model architecture). The reported SLME correlation of -0.76 is computed after training across the nine latent dimensions (Section III.B), and no latent parameter is optimized against SLME, so the correlation is an empirical observation rather than a fitted input renamed as a prediction. The discovery campaign (Section V, Discovery method) ranks materials by Manhattan distance from the top-SLME seed in the latent space and counts recovered top-20 SLME materials; the ranking is not derived from SLME labels, and the comparison against VAE, PCA, and random search is an empirical evaluation. This protocol does raise generalization concerns—there is no held-out split, only one distance metric, and one seed material—but those are robustness issues, not circularity. Self-citations to the DAE architecture [30, 37] introduce the model but do not carry the load of the new spectral application; the model is publicly implemented and independently published, so the citation is supporting evidence rather than an unverified premise. No equation or definition in the paper reduces a claimed prediction to its own input.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical entities. The main free parameters are the latent dimension and the disentanglement coefficients, chosen by hand. The key structural assumptions are that the DAE's orthogonality mechanism works as claimed, that SLME is a meaningful target, and that the latent-space distance search is a valid discovery heuristic.

free parameters (3)

latent_dim = 9
Chosen empirically as the smallest dimension allowing accurate reconstruction (Methods, Model architecture); a hand-tuned hyperparameter.
alpha = 0.005
DAE disentanglement hyperparameter, set by hand in Methods.
beta = 2
beta-VAE disentanglement hyperparameter, set by hand in Methods.

assumptions (4)

domain assumption The DAE's Euler layer enforces orthogonality of latent dimensions and yields disentangled representations
Invoked in Section III (see also ref [30]); if the orthogonality constraint does not hold as claimed, the interpretation of individual latent dimensions as independent factors is unsupported.
domain assumption SLME computed from the absorption spectrum and solar spectrum is a valid proxy for photovoltaic efficiency
SLME is maximized as the search target; if SLME is a poor proxy for real device efficiency, the discovery results and latent correlation lose their practical relevance.
domain assumption The optical absorption spectra from the two MPContribs datasets are sufficiently accurate and are unlabeled inputs to the autoencoder
All results depend on the quality and comparability of the precomputed spectra; normalization and dataset merging are not analyzed.
ad hoc to paper Manhattan distance to a single seed material in latent space is a meaningful ordering for photovoltaic potential
The discovery campaign (Section V, Methods) ranks candidates by Manhattan distance to the top-SLME material; no justification is given for why this ordering should track SLME, and the metric choice is not varied.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Learning disentangled latent representations facilitates discovery and design of functional materials." pith.science (2026). https://pith.science/paper/4W2UOCHW

@misc{pith2026250719602,
  author       = {Pith},
  title        = {Pith review of: Learning disentangled latent representations facilitates discovery and design of functional materials},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4W2UOCHW}},
  note         = {Machine review of arXiv:2507.19602}
}

read the original abstract

The discovery of new materials is often constrained by the need for large labelled datasets or expensive simulations. In this study, we explore the use of Disentangling Autoencoders (DAEs) to learn compact and interpretable representations of spectral data in an entirely unsupervised manner. We demonstrate that the DAE captures physically meaningful features in optical absorption spectra, relevant to photovoltaic (PV) performance, including a latent dimension strongly correlated with the Spectroscopic Limited Maximum Efficiency (SLME)--despite being trained without access to SLME labels. This feature corresponds to a well-known spectral signature: the transition from direct to indirect optical band gaps. Compared to Principal Component Analysis (PCA) and a beta-Variational Autoencoder (beta-VAE), the DAE achieves superior reconstruction fidelity, improved correlation with efficiency metrics, and more compact encoding of relevant features. We further show that the DAE latent space enables more efficient discovery of high-performing PV materials, identifying top candidates using fewer evaluations than both VAE-guided and random search. These results highlight the potential of DAEs as a powerful tool for unsupervised structure-property learning and suggest broad applicability to other areas of materials discovery where labeled data is limited but rich structure is present in raw signals.

Figures

Figures reproduced from arXiv: 2507.19602 by the authors.

**Figure 1.** FIG. 1. A depiction of the disentangled autoencoder used in this study. Residual layers and Euler layers are used to encode a [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** FIG. 2. Discovery rates for 9 latent dimensions using DAE [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** FIG. 3. Covariance matrix of maximum values in the spectrum and the SLME with respect to latent variables obtained using [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (3 more)

**Figure 4.** Figure 4: FIG. 4. Latent traversal of all latent dimensions using DAE. For each dimension the latent space is traversed from low latent [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]

**Figure 5.** Figure 5: FIG. 5. The zoomed-in view illustrates variations concerning [PITH_FULL_IMAGE:figures/full_fig_p006_5.png]

**Figure 6.** Figure 6: FIG. 6. Reconstruction comparison between DAE, VAE and PCA. The plots show the ground truth spectrum in blue and the [PITH_FULL_IMAGE:figures/full_fig_p007_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 35 canonical work pages

[1]

A. K. Cheetham, R. Seshadri, and F. Wudl, Chemical synthesis and materials discovery, Nature Synthesis 1, 514 (2022)

work page 2022
[2]

T. Hey, S. Tansley, K. M. Tolle, et al., The fourth paradigm: data-intensive scientific discovery, Vol. 1 (Mi- crosoft research Redmond, WA, 2009)

work page 2009
[3]

K. T. Butler, D. W. Davies, H. Cartwright, O. Isayev, and A. Walsh, Machine learning for molecular and ma- terials science, Nature559, 547 (2018)

2018
[4]

the projected expansion to 30–70 TW by 2050 de- mands the identification of novel, efficient absorber ma- terials. [5] Traditional approaches—based on experimen- tal synthesis and first-principles calculations—remain es- sential, but the increasing availability of large materials databases now presents new opportunities to accelerate discovery using machi...

work page Pith review arXiv 2025
[5]

IEA PVPS, Snapshot of global pv markets 2024 (2024)

work page 2024
[6]

N. M. Haegel, H. Atwater Jr, T. Barnes, C. Breyer, A. Burrell, Y.-M. Chiang, S. De Wolf, B. Dimmler, D. Feldman, S. Glunz, et al., Terawatt-scale photo- 8 voltaics: Transform global energy, Science 364, 836 (2019)

work page 2019
[7]

Scheidgen, L

M. Scheidgen, L. Himanen, A. N. Ladines, D. Sikter, M. Nakhaee, Á. Fekete, T. Chang, A. Golparvar, J. A. Márquez, S. Brockhauser,et al., Nomad: A distributed web-based platform for managing materials science re- search data, Journal of Open Source Software 8, 5388 (2023)

work page 2023
[8]

Kirklin, J

S. Kirklin, J. E. Saal, B. Meredig, A. Thompson, J. W. Doak, M. Aykol, S. Rühl, and C. Wolverton, The open quantum materials database (oqmd): assessing the accu- racy of dft formation energies, npj Computational Mate- rials 1, 1 (2015)

work page 2015

Show all 43 references

[9]

P.Ong, G.Hautier, W.Chen, W.D.Richards, S

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 approach to accelerating materials innovation, APL materials 1 (2013)

2013
[10]

Choudhary, K

K. Choudhary, K. F. Garrity, A. C. Reid, B. DeCost, A. J. Biacchi, 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 materials design, npj computa- t...

2020
[11]

S. P. Ong, W. D. Richards, A. Jain, G. Hautier, M. Kocher, S. Cholia, D. Gunter, V. L. Chevrier, K. A. Persson, and G. Ceder, Python materials genomics (py- matgen): A robust, open-source python library for mate- rials analysis, Computational Materials Science68, 314 (2013)

2013
[12]

A. H. Larsen, J. J. Mortensen, J. Blomqvist, I. E. Castelli, R. Christensen, M. Dułak, J. Friis, M. N. Groves, B. Hammer, C. Hargus,et al., The atomic sim- ulation environment—a python library for working with atoms, Journal of Physics: Condensed Matter29, 273002 (2017)

2017
[13]

Pizzi, A

G. Pizzi, A. Cepellotti, R. Sabatini, N. Marzari, and B. Kozinsky, Aiida: automated interactive infrastructure and database for computational science, Computational Materials Science 111, 218 (2016)

2016
[14]

Chibani and F.-X

S. Chibani and F.-X. Coudert, Machine learning ap- proaches for the prediction of materials properties, Apl Materials 8 (2020)

2020
[15]

R. Devi, K. T. Butler, and G. Sai Gautam, Optimal pre- train/fine-tune strategies for accurate material property predictions, npj Computational Materials10, 300 (2024)

2024
[16]

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

2016
[17]

L. M. Antunes, K. T. Butler, and R. Grau-Crespo, Pre- dicting thermoelectric transport properties from com- position with attention-based deep learning, Machine Learning: Science and Technology4, 015037 (2023)

2023
[18]

S.Kim, J.Noh, G.H.Gu, A.Aspuru-Guzik,andY.Jung, Generative adversarial networks for crystal structure pre- diction, ACS central science6, 1412 (2020)

2020
[19]

R. Jiao, W. Huang, P. Lin, J. Han, P. Chen, Y. Lu, and Y. Liu, Crystal structure prediction by joint equivari- ant diffusion, Advances in Neural Information Processing Systems 36, 17464 (2023)

2023
[20]

H. Park, Z. Li, and A. Walsh, Has generative artificial intelligence solved inverse materials design?, Matter 7, 2355 (2024)

2024
[21]

J. Noh, G. H. Gu, S. Kim, and Y. Jung, Machine-enabled inverse design of inorganic solid materials: promises and challenges, Chemical Science11, 4871 (2020)

2020
[22]

K. Wong, R. Qi, Y. Yang, Z. Luo, S. Guldin, and K. T. Butler, Predicting colloidal interaction parameters from small-angle x-ray scattering curves using artificial neural networks and markov chain monte carlo sampling, JACS Au 4, 3492 (2024)

2024
[23]

Mirza and K

A. Mirza and K. M. Jablonka, Elucidating structures from spectra using multimodal embeddings and discrete optimization, chemrXiv preprint 10.26434/chemrxiv- 2024-f3b18-v2 (2024)

2024 doi
[24]

Zhong, Interpretable machine learning: Fundamental principles and 10 grand challenges, Statistic Surveys16, 1 (2022)

C.Rudin, C.Chen, Z.Chen, H.Huang, L.Semenova,and C. Zhong, Interpretable machine learning: Fundamental principles and 10 grand challenges, Statistic Surveys16, 1 (2022)

2022
[25]

Oviedo, J

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

2022
[26]

Bengio, A

Y. Bengio, A. Courville, and P. Vincent, Representation learning: A review and new perspectives, IEEE trans- actions on pattern analysis and machine intelligence35, 1798 (2013)

2013
[27]

Spurek, A

P. Spurek, A. Nowak, J. Tabor, Ł. Maziarka, and S. Jas- trzębski, Non-linear ica based on cramer-wold metric, in International Conference on Neural Information Process- ing (Springer, 2020) pp. 294–305

2020
[28]

S. A. Eslami, D. Jimenez Rezende, F. Besse, F. Viola, A. S. Morcos, M. Garnelo, A. Ruderman, A. A. Rusu, I. Danihelka, K. Gregor,et al., Neural scene representa- tion and rendering, Science360, 1204 (2018)

2018
[29]

G. E. Hinton and R. R. Salakhutdinov, Reducing the dimensionality of data with neural networks, science313, 504 (2006)

2006
[30]

D. P. Kingma, Auto-encoding variational bayes, arXiv preprint arXiv:1312.6114 (2013)

2013 arXiv
[31]

Cha and J

J. Cha and J. Thiyagalingam, Orthogonality-enforced la- tent space in autoencoders: An approach to learning disentangled representations, inProceedings of the 40th International Conference on Machine Learning(PMLR,
[32]

Bridger, W

A. Bridger, W. I. David, T. J. Wood, M. Danaie, and K. T. Butler, Versatile domain mapping of scanning elec- tron nanobeam diffraction datasets utilising variational autoencoders, npj Computational Materials9, 14 (2023)

2023
[33]

Y. Baig, H. R. Ma, H. Xu, and L. You, Autoencoder neu- ral networks enable low dimensional structure analyses of microbial growth dynamics, Nature Communications14, 7937 (2023)

2023
[34]

Woods-Robinson, Y

R. Woods-Robinson, Y. Xiong, J.-X. Shen, N. Winner, M. K. Horton, M. Asta, A. M. Ganose, G. Hautier, and K. A. Persson, Designing transparent conductors using forbidden optical transitions, Matter6, 3021 (2023)

2023
[35]

D. H. Fabini, M. Koerner, and R. Seshadri, Candidate in- organic photovoltaic materials from electronic structure- based optical absorption and charge transport proxies, Chemistry of Materials, Chemistry of Materials31, 1561 (2019)

2019
[36]

Yu and A

L. Yu and A. Zunger, Identification of potential photo- voltaic absorbers based on first-principles<? format?> spectroscopic screening of materials, Physical review let- ters 108, 068701 (2012). 9

2012
[37]

Walker and K

M. Walker and K. T. Butler, The carbon cost of ma- terials discovery: Can machine learning really acceler- ate the discovery of new photovoltaics?, arXiv preprint arXiv:2507.13246 (2025)

2025 arXiv
[38]

J. Cha, J. Park, S. Pinilla, K. L. Morris, C. S. Allen, M. I. Wilkinson, and J. Thiyagalingam, Discovering fully semantic representations via centroid-and orientation- aware feature learning, Nature Machine Intelligence , 1 (2025)

2025
[39]

C. P. Burgess, I. Higgins, A. Pal, L. Matthey, N. Watters, G. Desjardins, and A. Lerchner, Understanding disentan- gling in β-vae, arXiv preprint arXiv:1804.03599 (2018)

2018 arXiv
[40]

P. Huck, A. Jain, D. Gunter, D. Winston, and K. Pers- son, A community contribution framework for sharing materials data with materials project, in2015 IEEE 11th International Conference on e-Science(IEEE, 2015) pp. 535–541

2015
[41]

P.Huck, A.Jain, D.Gunter, D.Winston,andK.Persson, A community contribution framework for sharing mate- rials data with materials project, in2015 IEEE 11th In- ternational Conference on e-Science(2015) pp. 535–541

2015
[42]

P. Huck, D. Gunter, S. Cholia, D. Winston, A. T. N’Diaye, and K. Persson, User applications driven by the community contribution framework mpcontribs in the materials project, Concurrency and Com- putation: Practice and Experience 28, 1982 (2016), https://onlinelibrary.wiley.c...

2016 doi
[43]

Code available at: https://github.com/ Jaehoon-Cha-Data/material_discovery/tree/main

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.

[1] [1]

A. K. Cheetham, R. Seshadri, and F. Wudl, Chemical synthesis and materials discovery, Nature Synthesis 1, 514 (2022)

work page 2022

[2] [2]

T. Hey, S. Tansley, K. M. Tolle, et al., The fourth paradigm: data-intensive scientific discovery, Vol. 1 (Mi- crosoft research Redmond, WA, 2009)

work page 2009

[3] [3]

K. T. Butler, D. W. Davies, H. Cartwright, O. Isayev, and A. Walsh, Machine learning for molecular and ma- terials science, Nature559, 547 (2018)

2018

[4] [4]

the projected expansion to 30–70 TW by 2050 de- mands the identification of novel, efficient absorber ma- terials. [5] Traditional approaches—based on experimen- tal synthesis and first-principles calculations—remain es- sential, but the increasing availability of large materials databases now presents new opportunities to accelerate discovery using machi...

work page Pith review arXiv 2025

[5] [5]

IEA PVPS, Snapshot of global pv markets 2024 (2024)

work page 2024

[6] [6]

N. M. Haegel, H. Atwater Jr, T. Barnes, C. Breyer, A. Burrell, Y.-M. Chiang, S. De Wolf, B. Dimmler, D. Feldman, S. Glunz, et al., Terawatt-scale photo- 8 voltaics: Transform global energy, Science 364, 836 (2019)

work page 2019

[7] [7]

Scheidgen, L

M. Scheidgen, L. Himanen, A. N. Ladines, D. Sikter, M. Nakhaee, Á. Fekete, T. Chang, A. Golparvar, J. A. Márquez, S. Brockhauser,et al., Nomad: A distributed web-based platform for managing materials science re- search data, Journal of Open Source Software 8, 5388 (2023)

work page 2023

[8] [8]

Kirklin, J

S. Kirklin, J. E. Saal, B. Meredig, A. Thompson, J. W. Doak, M. Aykol, S. Rühl, and C. Wolverton, The open quantum materials database (oqmd): assessing the accu- racy of dft formation energies, npj Computational Mate- rials 1, 1 (2015)

work page 2015