Pith. sign in

REVIEW 3 major objections 6 minor 43 references

SynCoTrain: A Dual Classifier PU-learning Framework for Synthesizability Prediction

T0 review · 3 major / 6 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read SynCoTrain claims a co-training pair of graph neural networks, using positive-and-unlabeled learning, can label theoretical oxide crystals as synthesizable or not with 95–97% recall, without any negative training examples.

desk verdict Co-training is a plausible new combination, but the ground-truth validation is calibrated to the method and doesn't yet support the reliability claim. read the letter →

arxiv 2411.12011 v1 pith:LKT3AGLA submitted 2024-11-18 cond-mat.mtrl-sci cs.LG

Sasan Amariamir , Janine George , Philipp Benner This is my paper · ORCID
classification cond-mat.mtrl-scics.LG
keywords synthesizabilitypredictionpositiveandunlabeledlearningco-traininggraphneuralnetworksSchNetALIGNNoxidecrystalsmaterialsdiscovery
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

SynCoTrain is a machine-learning method for predicting whether a proposed oxide crystal can actually be synthesized, using only positive examples of known materials plus a large pool of unlabeled theoretical crystals. The paper argues that by alternating two structurally different graph neural networks in a co-training loop and treating the unlabeled set through positive-and-unlabeled (PU) learning, the method can label theoretical crystals reliably without any explicit negative data. On oxide data from the Materials Project, the method reports 95–97% recall on known positives and labels about 21% of the unlabeled theoretical crystals as synthesizable. The authors also construct a separate stability-classification task with true labels and show that the PU recall approximates the ground-truth recall, which they take as evidence that the reported synthesizability recall is trustworthy. If correct, the model offers a practical filter for high-throughput materials discovery and generative design.

What carries the argument

The central machinery is co-training between two structurally distinct graph neural networks, SchNet and ALIGNN, each wrapped in a bagging PU learner. At each iteration, one network trains on the current positive set plus a random subset of unlabeled data treated as negative; it then scores the rest of the unlabeled data, and confident positives are added to the positive class for the other network. Two mirrored series of iterations run with opposite starting networks, and after the optimal iteration their scores are averaged with a 0.5 cutoff to assign final synthesizability labels. These labels are then used to train a final SchNet classifier for prediction on new crystals.

What would settle it

Attempted-synthesis data for a set of theoretical oxide crystals (or a fully labeled subset) would settle the claim: if SynCoTrain's recall on truly synthesizable examples is substantially below the reported 95–97%, the method's headline reliability fails. A cheaper check is to rerun the stability ground-truth experiment with the conventional 0.1 eV threshold: the paper notes that threshold yields low recall; if that recall also no longer tracks ground truth, the validation is specific to the unusual 0.015 eV cutoff.

Watch

Extended reading notes

Core claim

SynCoTrain claims that synthesizability of oxide crystals can be reliably predicted without negative data by co-training two complementary graph convolutional networks, SchNet and ALIGNN, as alternating base PU learners. Starting from experimental crystals as the positive class, each iteration expands the positive class with confidently predicted positives from the other model, and the final synthesizability score is the average of the two series' scores. On the Materials Project oxide dataset, the resulting labels achieve a recall range of 95–97% on known positives while only 21% of the unlabeled data are marked synthesizable. The paper validates this recall by a ground-truth experiment: a stability-prediction PU task where true labels are known shows that the recall estimates track the true recall across co-training iterations. The authors further show that the synthesizability scores correlate with energy above the convex hull—over 99% of predicted-synthesizable crystals sit below 1 eV above the hull—but that stability alone is not enough, since only about 21% of crystals below that hull distance are classified synthesizable.

Load-bearing premise

The reliability demonstration depends on treating 'stable' and 'unstable' labels assigned at an energy threshold of 0.015 eV above the convex hull—a cutoff near the precision limit of density-functional-theory calculations—as reliable ground truth; if those labels are not meaningful, the evidence that the PU recall approximates real performance on unlabeled crystals would collapse.

Editorial extensions

If this is right

  • A model that filters out a large fraction of unsynthesizable hypothetical oxides before expensive DFT screening could save significant compute and experiment time in high-throughput discovery pipelines.
  • The same co-training procedure can be applied to other material families or to compounds beyond oxides, provided a positive-only dataset and a large unlabeled pool are available.
  • Because the final labels are produced by averaging two independently biased networks, the approach offers a way to reduce single-model bias in materials-property classification.
  • The 21% predicted-synthesizable fraction among unlabeled theoretical oxides suggests most hypothetical oxides are unlikely to be synthesizable, which could inform how generative models propose new candidates.

Reading between the lines

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

  • The paper's own threshold analysis implies that the 0.5 decision cutoff is tunable: a stricter cutoff could be used for resource-constrained screening, and a looser one for exploratory generative design; the paper illustrates the trade-off but does not formalize an optimal choice.
  • A natural extension beyond the paper is to test whether the co-trained labels improve downstream generative models, for example by feeding SynCoTrain's filter into inverse-design loops to reduce the rate of unsynthesizable proposals.
  • Because the ground-truth stability validation uses a stability proxy, the method's true generalization to kinetic or technological synthesizability remains untested; an experimental verification on a small set of novel oxides would be a direct test of the synthesizability labels.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The manuscript introduces SynCoTrain, a co-training framework that combines two graph neural networks (SchNet and ALIGNN) with positive-unlabeled (PU) learning to predict the synthesizability of oxide crystals. Starting from experimental oxides as known positives and theoretical oxides as unlabeled data, the method iteratively expands the positive class by exchanging confident predictions between the two classifiers, averages the scores from two mirrored training series, and uses the resulting labels to train a final SchNet predictor. The central claims are that the PU recall on labeled positives is reliably high (a 'recall range' of 95-97%), that 21% of the unlabeled theoretical oxides are predicted synthesizable, and that the final predictor achieves 90.5% accuracy. The authors also present a 'ground truth' stability-control experiment intended to show that their PU recall approximates true recall.

Significance. If its claims were established, SynCoTrain would be a practically useful filter for high-throughput materials discovery, and the combination of co-training with PU learning is a sensible response to the scarcity of negative synthesis data. The paper ships its code and data, which is commendable. The stability-control experiment is a creative attempt to address a real weakness of PU evaluation (that recall on known positives may not reflect recall on all positives). However, as detailed in the major comments, the experiment is undermined by the post hoc choice of a near-precision-limit stability threshold, and the final model is evaluated circularly against its own generated labels, so the paper currently does not deliver on its reliability claim.

major comments (3)
  1. [§2.3.2 and Supplemental Material] The ground-truth stability experiment uses a stability threshold of 0.015 eV above the convex hull, a value chosen explicitly because the conventional 0.1 eV threshold 'would not work well for our demonstration' (Supplemental Material). At 0.015 eV the labels are near the precision limits of DFT, and the supplemental analysis shows that 97% of the 6,602 misclassified points are false positives (predicted stable but labeled unstable), with 85% of these errors lying below 0.1 eV above hull. This pattern is exactly what one would expect from a model learning a threshold at the numerical noise floor, rather than learning a physically meaningful notion of stability. The experiment therefore does not establish that the PU recall on labeled positives approximates the recall on all positives, and without that transfer the reported [95-97]% recall and the 21% positive rate on unlabeled theoretical oxides are compatible with low precision on that set. The claim in the abstract and in §2.5 that SynCoTrain is a 'reliable tool for predicting synthesizability' is not supported by this evidence.
  2. [§2.4 and §3.4] The final synthesizability predictor is trained on labels produced by the co-training process itself, and the reported 90.5% accuracy is measured on a test set taken from the same data with those same generated labels. This is a circular evaluation: it measures the predictor's agreement with the co-training output, not with any external ground truth for synthesizability. The paper does not validate the final labels against independent evidence, such as recently reported oxide syntheses or a carefully curated set of known unsynthesizable structures. Moreover, the use of accuracy as the metric contradicts the paper's own argument in §2.2 that accuracy is inappropriate in PU settings because it requires negative labels; the only reason accuracy can be computed here is that the negative labels were manufactured by the same model that is being evaluated. The 90.5% figure therefore does not support the conclusion that SynCoTrain is a reliable synthesizability predictor, and the reader is left without any externally grounded measure of the final model's performance.
  3. [§2.2 and §3.1] The paper does not demonstrate that co-training provides a benefit over its constituent single-model PU learners. The recall increases from iteration '0' to iteration '2' shown in Fig. 2 are accompanied by a growing positive class (the pseudo-positives added at each iteration), so part or all of the improvement could be due simply to having more training data rather than to the collaborative exchange between the two classifiers. A control in which a single PU learner is trained on the same expanded positive set, without the alternating two-series exchange, is needed to isolate the effect of co-training. Without such a control, the central methodological claim that co-training 'mitigates model bias and enhances generalizability' (§1) is not supported by the evidence presented.
minor comments (6)
  1. [§2.2] The sentence 'The construction and reasoning behind this are detailed in the Ground Truth evaluation section.' appears twice verbatim; one occurrence should be removed.
  2. [Supplemental Material] Two citations are left as placeholder '[?]' after 'contamination' and 'Δ-machine learning'; these need to be completed before publication.
  3. [§3.4 vs. §2.2] The final predictor is trained on labels assigned with a 0.75 cutoff (§3.4), while the co-training final labels and the reported 21% positive rate are based on a 0.5 cutoff (§2.2). The paper should clarify which label set is used to compute the 90.5% accuracy and how the two cutoffs relate.
  4. [§2.2 and Fig. 6] The statement that unstable crystals are '2.5 times less likely' to be classified synthesizable should report the underlying fractions (e.g., the percentage labeled synthesizable among crystals above 1 eV above hull vs. below that threshold).
  5. [§2.2] The term 'recall range' is used throughout but not formally defined in the main text; it should be defined as the interval between the dynamic test-set recall and the leave-out recall, and the paper should state which of the two values is the upper bound.
  6. [§2.4] Since the paper argues in §2.2 that accuracy is not an appropriate metric for PU evaluation, the accuracy reported for the final predictor in §2.4 should be accompanied by a note explaining that it is computed against the co-training-generated labels and therefore is not a ground-truth measure.

Circularity Check

2 steps flagged · score 5.0 of 10

Reliability claim rests on a tailored stability benchmark and a final predictor evaluated on its own co-training labels; recall on experimental positives is genuinely external but does not by itself transfer to unlabeled crystals.

  1. self definitional [Section 2.4 'Predicting Synthesizability' and Section 3.4 'The synthesizability predictor']
    "Once we have synthesizability labels for both the experimental and theoretical data, a simple machine learning task remains. We train a classifier on these labels and end up with a model that can predict synthesizability... The trained model reached 90.5% accuracy on a test set comprising 5,180 data points."

    The 'synthesizability labels' used to train the final classifier are not external ground truth; they are the averaged co-training scores from iteration '2', thresholded at 0.75 or 0.5. The test set is a holdout split of those same generated labels. Thus the 90.5% accuracy measures how well the final SchNet model reproduces the co-training ensemble's own labels, not how well it predicts synthesizability. Any claim that this accuracy validates a 'reliable tool for predicting synthesizability' is self-referential unless the generated labels themselves are externally validated, and the paper's only external validation attempt is the stability demonstration with a threshold chosen to make the method look good.

  2. fitted input called prediction [Supplemental Material, 'Ground truth stability set-up']
    "While the obvious choice would have been 0.1 eV, a commonly used threshold for stability, it would not work well for our demonstration. ... Instead, we chose 0.015 eV as the threshold for stability, which labels approximately a quarter (26%) of our data as stable. This proportion aligns better with what we expect for synthesizability and provides a more suitable demonstration of the model’s capabilities."

    The only evidence that PU recall on known positives approximates recall on all positives comes from this stability experiment. The stability threshold is not an independent, physically standard choice: the paper explicitly rejects 0.1 eV because recall would be 'quite low' and a 'poor demonstration', and instead selects 0.015 eV to yield a 26% positive rate that matches the expected synthesizability rate and the PU algorithm's low-contamination assumption. The validation target is therefore fitted to the method's desired operating regime, and the observed agreement between test-set recall and ground-truth recall is a consequence of this construction rather than independent evidence that 95-97% recall on experimental oxides transfers to the unlabeled synthesizability task.

full rationale

The paper does contain genuinely external components: recall on the known experimental positives is a real measured quantity, and the correlation of synthesizability scores with energy above the convex hull (Fig. 6/7) is an independent sanity check that the model has learned stability-related signal. However, the central reliability claim depends on a chain that is only partially grounded. The recall statistics themselves are computed on experimental positive subsets, which is legitimate, but the paper's only bridge from those known-positive recalls to recall over all positives is the stability ground-truth experiment where the threshold is explicitly chosen to avoid a 'poor demonstration' and to produce a desired 26% positive rate. That makes the key validation a fitted input rather than an externally falsifiable benchmark. Additionally, the final synthesizability predictor is trained on labels generated by the co-training process itself, and its reported 90.5% accuracy is on a test set drawn from those same generated labels, so this number is a self-referential fidelity check. The external OQMD, WBM, and iMatGen datasets are used only to show score distributions, not to validate correctness. Self-citations are not load-bearing, and there is no imported uniqueness theorem. On balance, this is partial circularity: the core claim is not forced by definition because real experimental-positive recall and stability correlations provide independent content, but the load-bearing reliability evidence is tailored and the final predictor evaluation is self-referential.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central claim rests on several hand-chosen thresholds and unverified domain assumptions. The most significant are the 0.015 eV ground-truth stability threshold (chosen post hoc), the arbitrary 0.5/0.75 decision thresholds, and the assumed conditional independence of the two GNN views for co-training. No new physical entities are introduced.

free parameters (6)
  • Ground-truth stability threshold (energy above hull) = 0.015 eV
    Chosen in the supplemental ground-truth experiment because the conventional 0.1 eV threshold would produce low recall and a poor demonstration; the 0.015 eV threshold yields 26% stable labels, matching the expected synthesizability proportion. Selection is post hoc and near DFT precision limits.
  • Optimal co-training iteration = 2
    Selected based on recall progression and predicted positive rate on test sets; iteration 3 showed diminishing returns. Choosing the iteration on the evaluation sets risks mild overfitting to those sets.
  • Class expansion threshold = 0.75
    Unlabeled points with averaged synthesizability score above 0.75 are added to the positive class; the paper acknowledges thresholds are arbitrary and adjustable.
  • Labeling threshold = 0.5
    Used to assign final synthesizability labels after averaging scores. Called 'unbiased' but acknowledged as arbitrary.
  • Label noise fraction = 5% of each class
    Randomly flips 5% of positive labels to 0 and same number of negative to 1 in final predictor training, a regularization choice.
  • Loss weighting ratio = 0.45:0.55 (positive:negative)
    Adjusted to discourage over-prediction of the positive class in the final synthesizability predictor.
assumptions (6)
  • domain assumption The unlabeled data are mostly negative, i.e., low contamination by positives, as required by the bagging PU learning method.
    Invoked in Methods 3.2 ('PU learning performs best when this contamination is low') and foundational to Mordelet and Vert. The paper later finds 21% of unlabeled data predicted positive, which is moderate contamination.
  • ad hoc to paper SchNet and ALIGNN provide conditionally independent views of the data given the label, a requirement for co-training to converge to the correct target.
    Section 3.1 states the two encodings provide different views and are complementary, but no test of conditional independence is provided; co-training theory (Blum and Mitchell) requires this.
  • domain assumption All experimental ICSD oxides are synthesizable and can serve as true positives.
    Used throughout to define the positive class and to compute recall; the paper does not audit ICSD entries for synthesis conditions or context.
  • domain assumption DFT energies from Materials Project (version 2023.11.1) are accurate enough to support a 0.015 eV stability threshold.
    The supplemental ground-truth experiment labels materials stable/unstable at 0.015 eV above hull, close to DFT precision limits; the paper acknowledges this concern.
  • ad hoc to paper Adding Gaussian noise to atomic positions (data augmentation) preserves synthesizability labels.
    Used in final predictor training; assumes perturbed structures remain valid members of the same class.
  • domain assumption Experimental data with energy above hull greater than 1 eV (less than 1% of data) are corrupt and can be removed.
    Section 2.1: 'Less than 1% of the experimental data with energy above hull higher than 1eV were removed, as potentially corrupt data.'

how reviews work

0 comments
Cite this review

Pith. "Pith review of SynCoTrain: A Dual Classifier PU-learning Framework for Synthesizability Prediction." pith.science (2026). https://pith.science/paper/LKT3AGLA

@misc{pith2026241112011,
  author       = {Pith},
  title        = {Pith review of: SynCoTrain: A Dual Classifier PU-learning Framework for Synthesizability Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LKT3AGLA}},
  note         = {Machine review of arXiv:2411.12011}
}
read the original abstract

Material discovery is a cornerstone of modern science, driving advancements in diverse disciplines from biomedical technology to climate solutions. Predicting synthesizability, a critical factor in realizing novel materials, remains a complex challenge due to the limitations of traditional heuristics and thermodynamic proxies. While stability metrics such as formation energy offer partial insights, they fail to account for kinetic factors and technological constraints that influence synthesis outcomes. These challenges are further compounded by the scarcity of negative data, as failed synthesis attempts are often unpublished or context-specific. We present SynCoTrain, a semi-supervised machine learning model designed to predict the synthesizability of materials. SynCoTrain employs a co-training framework leveraging two complementary graph convolutional neural networks: SchNet and ALIGNN. By iteratively exchanging predictions between classifiers, SynCoTrain mitigates model bias and enhances generalizability. Our approach uses Positive and Unlabeled (PU) Learning to address the absence of explicit negative data, iteratively refining predictions through collaborative learning. The model demonstrates robust performance, achieving high recall on internal and leave-out test sets. By focusing on oxide crystals, a well-characterized material family with extensive experimental data, we establish SynCoTrain as a reliable tool for predicting synthesizability while balancing dataset variability and computational efficiency. This work highlights the potential of co-training to advance high-throughput materials discovery and generative research, offering a scalable solution to the challenge of synthesizability prediction.

Figures

Figures reproduced from arXiv: 2411.12011 by the authors.

Figure 1
Figure 1. Overview of the Workflow in SynCoTrain a) The PU data is passed to two distinct PU classifiers, each learning from a different view of the data. Each classifier labels unlabeled data points as positive or negative. The new labels from each PU classifier are used to expand the positive class for retraining the other classifier. b) After co-training steps, each unlabeled data point receives a prediction score from eac… view at source ↗
Figure 2
Figure 2. Recall progression per iteration for both co-training series 12 [PITH_FULL_IMAGE:figures/full_fig_p012_2.png] view at source ↗
Figure 3
Figure 3. Synthesizability score distribution for Iteration ’0’ for the a) ALIGNN0 series and b) SchNet0 series. the predicted positive rate increases in both series for iteration ‘3’, without a meaningful increase in recall range to justify it. This means that the model is more likely to classify a theoretical crystal as synthesizable, without improving its understanding of synthesizability. This is akin to over-fitting, whe… view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Synthesizability score distribution for Iteration ’2’ for the a) ALIGNN0 series and b) SchNet0 series. 20% of the unlabeled data as synthesizable. The synthesizability scores for the intermediate iterations can be found in the supplementary materials. In the final step…
Figure 5
Figure 5. Figure 5: Label distribution after averaging scores 15 [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]
Figure 6
Figure 6. Figure 6: Density scatter plot of energy above hull vs the synthesizability score for the unlabeled data. The logarithmic color bar on the side indicates density map of the unlabeled crystals that are classified as synthesizable, have an energy less than 1eV above the convex hul…
Figure 7
Figure 7. Figure 7: Distribution of formation energy versus energy above hull for experimental data (left) and theoretical data (right), separated by predicted labels. The figure illustrates the expected clustering of positively labeled data around lower values of energy above hull while …
Figure 8
Figure 8. Figure 8: Ground Truth Recall progression per iteration for stability classe. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_8.png]
Figure 9
Figure 9. Figure 9: Synthesizability probability distributions across theoretical databases, expressed as a percentage of each dataset. The first distribution represents the theoretical portion of the test set, selected from Materials Project data. The second distribution corresponds to v…
Figure 10
Figure 10. Figure 10: Label distribution based on 0.25 and 0.75 classification thresholds at the end of co-training. In this work we combined two different learners based on strong classifiers to reach a 25 [PITH_FULL_IMAGE:figures/full_fig_p025_10.png]
Figure 1
Figure 1. Figure 1: Synthesizability score distribution for Iteration ’0’ for the a) ALIGNN0 series and b) SchNet0 series. (a) (b) [PITH_FULL_IMAGE:figures/full_fig_p037_1.png]
Figure 2
Figure 2. Figure 2: Synthesizability score distribution for Iteration ’1’ for the a) ALIGNN0 series and b) SchNet0 series. 1 arXiv:2411.12011v1 [cond-mat.mtrl-sci] 18 Nov 2024 [PITH_FULL_IMAGE:figures/full_fig_p037_2.png]
Figure 3
Figure 3. Figure 3: Synthesizability score distribution for Iteration ’2’ for the a) ALIGNN0 series and b) SchNet0 series. (a) (b) [PITH_FULL_IMAGE:figures/full_fig_p038_3.png]
Figure 4
Figure 4. Figure 4: Synthesizability score distribution for Iteration ’3’ for the a) ALIGNN0 series and b) SchNet0 series. 2 [PITH_FULL_IMAGE:figures/full_fig_p038_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

43 extracted references · 42 canonical work pages

  1. [1]

    J., Jackson, N

    De Pablo, J. J., Jackson, N. E., Webb, M. A., Chen, L.-Q., Moore, J. E., Morgan, D., Jacobs, R., Pollock, T., Schlom, D. G., Toberer, E. S., Analytis, J., Dabo, I., DeLongchamp, D. M., Fiete, G. A., Grason, G. M., Hautier, G., Mo, Y., Rajan, K., Reed, E. J., Rodriguez, E., Stevanovic, V., Suntivich, J., Thornton, K., and Zhao, J.-C. (April, 2019) New fron...

    work page 2019

  2. [2]

    (August, 2012) The Materials Genome Initiative: One year on

    White, A. (August, 2012) The Materials Genome Initiative: One year on. MRS Bulletin, 37(8), 715–716

    work page 2012

  3. [3]

    Rodgers, J. R. and Cebon, D. (December, 2006) Materials Informatics. MRS Bulletin, 31(12), 975–980

    work page 2006

  4. [4]

    (April, 1929) THE PRINCIPLES DETERMINING THE STRUCTURE OF COMPLEX IONIC CRYSTALS

    Pauling, L. (April, 1929) THE PRINCIPLES DETERMINING THE STRUCTURE OF COMPLEX IONIC CRYSTALS. Journal of the American Chemical Society, 51(4), 1010–1026 Publisher: American Chemical Society

    work page 1929

  5. [5]

    preprint, In Review (March, 2023)

    Antoniuk, E., Cheon, G., Wang, G., Bernstein, D., Cai, W., and Reed, E., Predicting the Synthesizability of Crystalline Inorganic Materials from the Data of Known Material 30 Compositions. preprint, In Review (March, 2023)

    work page 2023

  6. [6]

    P., Hautier, G., Chen, W., Richards, W

    Jain, A., Ong, S. P., Hautier, G., Chen, W., Richards, W. D., Dacek, S., Cholia, S., Gunter, D., Skinner, D., Ceder, G., and Persson, K. A. (July, 2013) Commentary: The Materials Project: A materials genome approach to accelerating materials innovation. APL Materials, 1(1), 011002 Publisher: American Institute of Physics

    work page 2013

  7. [7]

    (May, 2020) The Limited Predictive Power of the Pauling Rules

    George, J., Waroquiers, D., Di Stefano, D., Petretto, G., Rignanese, G., and Hautier, G. (May, 2020) The Limited Predictive Power of the Pauling Rules. Angewandte Chemie, 132(19), 7639–7645

    work page 2020

  8. [8]

    T., Ong, S

    Sun, W., Dacek, S. T., Ong, S. P., Hautier, G., Jain, A., Richards, W. D., Gamst, A. C., Persson, K. A., and Ceder, G. (November, 2016) The thermodynamic scale of inorganic crystalline metastability. Science Advances, 2(11), e1600225

    work page 2016

Show all 43 references
  1. [9]

    Cerqueira, T. F. T., Lin, S., Amsler, M., Goedecker, S., Botti, S., and Marques, M. A. L. (July, 2015) Identification of Novel Cu, Ag, and Au Ternary Oxides from Global Structural Prediction. Chemistry of Materials, 27(13), 4562–4573 Publisher: American Chemical Society

    2015

  2. [10]

    K., Montoya, J

    Singh, A. K., Montoya, J. H., Gregoire, J. M., and Persson, K. A. (January, 2019) Robust and synthesizable photocatalysts for CO2 reduction: a data-driven materials discovery. Nature Communications, 10(1), 443

    2019

  3. [11]

    S., Sanchez-Lengeling, B., Gregoire, J

    Noh, J., Kim, J., Stein, H. S., Sanchez-Lengeling, B., Gregoire, J. M., Aspuru-Guzik, A., and Jung, Y. (November, 2019) Inverse Design of Solid-State Materials via a Continuous Representation. Matter, 1(5), 1370–1384. 31

    2019

  4. [12]

    (December, 2012) First princi- ples high throughput screening of oxynitrides for water-splitting photocatalysts

    Wu, Y., Lazic, P., Hautier, G., Persson, K., and Ceder, G. (December, 2012) First princi- ples high throughput screening of oxynitrides for water-splitting photocatalysts. Energy & Environmental Science, 6(1), 157–168 Publisher: The Royal Society of Chemistry

    2012

  5. [13]

    Bartel, C. J. (February, 2022) Review of computational approaches to predict the thermodynamic stability of inorganic solids. Journal of Materials Science,

    2022

  6. [14]

    H., Noh, J., Kim, J., and Jung, Y

    Jang, J., Gu, G. H., Noh, J., Kim, J., and Jung, Y. (November, 2020) Structure-Based Synthesizability Prediction of Crystals Using Partially Supervised Learning. Journal of the American Chemical Society, 142(44), 18836–18843

    2020

  7. [15]

    E., Ward, L., Borg, C., Mehta, A., and Wolverton, C

    Lee, A., Sarker, S., Saal, J. E., Ward, L., Borg, C., Mehta, A., and Wolverton, C. (October, 2022) Machine learned synthesizability predictions aided by density functional theory. Communications Materials, 3(1), 73

    2022

  8. [16]

    and Chen, W

    Li, K. and Chen, W. (June, 2021) Recent progress in high-entropy alloys for catalysts: synthesis, applications, and prospects. Materials Today Energy, 20, 100638

    2021

  9. [17]

    (October, 2020) Chemistry under high pressure

    Miao, M., Sun, Y., Zurek, E., and Lin, H. (October, 2020) Chemistry under high pressure. Nature Reviews Chemistry, 4(10), 508–527 Publisher: Nature Publishing Group

    2020

  10. [18]

    J., Needs, R

    Liu, C., Gao, H., Hermann, A., Wang, Y., Miao, M., Pickard, C. J., Needs, R. J., Wang, H.-T., Xing, D., and Sun, J. (April, 2020) Plastic and Superionic Helium Ammonia Compounds under High Pressure and High Temperature. Physical Review X, 10(2), 021007

    2020

  11. [19]

    K., and Mingo, N

    Legrain, F., Carrete, J., van Roekeghem, A., Madsen, G. K., and Mingo, N. (January,

  12. [20]

    (July, 2019) Semi-supervised machine-learning classification of materials synthesis procedures

    Huo, H., Rong, Z., Kononova, O., Sun, W., Botari, T., He, T., Tshitoyan, V., and Ceder, G. (July, 2019) Semi-supervised machine-learning classification of materials synthesis procedures. npj Computational Materials, 5(1), 62

    2019

  13. [21]

    (December, 2021) Predicting synthesizability of crystalline materials via deep learning

    Davariashtiyani, A., Kadkhodaie, Z., and Kadkhodaei, S. (December, 2021) Predicting synthesizability of crystalline materials via deep learning. Communications Materials, 2(1), 115

    2021

  14. [22]

    I., Hung, L., Suram, S., Herring, P., Wolverton, C., and Hum- melshøj, J

    Aykol, M., Hegde, V. I., Hung, L., Suram, S., Herring, P., Wolverton, C., and Hum- melshøj, J. S. (May, 2019) Network analysis of synthesizable materials discovery. Nature Communications, 10(1), 2018

    2019

  15. [23]

    H., Jang, J., Noh, J., Walsh, A., and Jung, Y

    Gu, G. H., Jang, J., Noh, J., Walsh, A., and Jung, Y. (April, 2022) Perovskite synthesiz- ability using graph neural networks. npj Computational Materials, 8(1), 71

    2022

  16. [24]

    C., Adler, P

    Raccuglia, P., Elbert, K. C., Adler, P. D. F., Falk, C., Wenny, M. B., Mollo, A., Zeller, M., Friedler, S. A., Schrier, J., and Norquist, A. J. (May, 2016) Machine-learning-assisted materials discovery using failed experiments. Nature, 533(7601), 73–76

    2016

  17. [25]

    Li, K., DeCost, B., Choudhary, K., Greenwood, M., and Hattrick-Simpers, J. (April,

  18. [26]

    and Mitchell, T

    Blum, A. and Mitchell, T. (July, 1998) Combining labeled and unlabeled data with co-training. In Proceedings of the eleventh annual conference on Computational learning theory Madison Wisconsin USA: ACM pp. 92–100

    1998

  19. [27]

    (2003) Text Classification and Co-training from Positive and Unlabeled Examples

    Denis, F., Laurent, A., Gilleron, R., and Tommasi, M. (2003) Text Classification and Co-training from Positive and Unlabeled Examples. Proceedings of the ICML 2003 workshop: the continuum from labeled to unlabeled data, p. 8

    2003

  20. [28]

    and DeCost, B

    Choudhary, K. and DeCost, B. (November, 2021) Atomistic Line Graph Neural Network for improved materials property predictions. npj Computational Materials, 7(1), 185

    2021

  21. [29]

    T., Kessel, P., Gastegger, M., Nicoli, K

    Sch¨ utt, K. T., Kessel, P., Gastegger, M., Nicoli, K. A., Tkatchenko, A., and M¨ uller, K.-R. (January, 2019) SchNetPack: A Deep Learning Toolbox For Atomistic Systems. Journal of Chemical Theory and Computation, 15(1), 448–455

    2019

  22. [30]

    T., Kindermans, P.-J., Sauceda, H

    Sch¨ utt, K. T., Kindermans, P.-J., Sauceda, H. E., Chmiela, S., Tkatchenko, A., and M¨ uller, K.-R. (December, 2017) SchNet: A continuous-filter convolutional neural network for modeling quantum interactions. arXiv:1706.08566 [physics, stat], arXiv: 1706.08566

    2017 arXiv

  23. [31]

    and Vert, J.-P

    Mordelet, F. and Vert, J.-P. (February, 2014) A bagging SVM to learn from positive and unlabeled examples. Pattern Recognition Letters, 37, 201–209

    2014

  24. [32]

    and Khomenkova, L

    Savkina, R. and Khomenkova, L. (May, 2020) Oxide-Based Materials and Structures: Fundamentals and Applications, CRC Press, Google-Books-ID: Q7XjDwAAQBAJ

    2020

  25. [33]

    (Octo- 34 ber, 2017) Statistical Analysis of Coordination Environments in Oxides

    Waroquiers, D., Gonze, X., Rignanese, G.-M., Welker-Nieuwoudt, C., Rosowski, F., G¨ obel, M., Schenk, S., Degelmann, P., Andr´ e, R., Glaum, R., and Hautier, G. (Octo- 34 ber, 2017) Statistical Analysis of Coordination Environments in Oxides. Chemistry of Materials, 29(19), 8346–8360

    2017

  26. [34]

    (2004) The Inorganic Crystal Structure Database (ICSD)—Present and Future

    Hellenbrandt, M. (2004) The Inorganic Crystal Structure Database (ICSD)—Present and Future. Crystallography Reviews, 10(1), 17–22

    2004

  27. [35]

    P., Richards, W

    Ong, S. P., Richards, W. D., Jain, A., Hautier, G., Kocher, M., Cholia, S., Gunter, D., Chevrier, V. L., Persson, K. A., and Ceder, G. (February, 2013) Python Materials Genomics (pymatgen): A robust, open-source python library for materials analysis. Computational Materials Sc...

    2013

  28. [36]

    E., Meredig, B., Thompson, A., Doak, J

    Kirklin, S., Saal, J. E., Meredig, B., Thompson, A., Doak, J. W., Aykol, M., R¨ uhl, S., and Wolverton, C. (December, 2015) The Open Quantum Materials Database (OQMD): assessing the accuracy of DFT formation energies. npj Computational Materials, 1(1), 15010

    2015

  29. [37]

    F., Reid, A

    Choudhary, K., Garrity, K. F., Reid, A. C. E., DeCost, B., Biacchi, A. J., Hight Walker, A. R., Trautt, Z., Hattrick-Simpers, J., Kusne, A. G., Centrone, A., Davydov, A., Jiang, J., Pachter, R., Cheon, G., Reed, E., Agrawal, A., Qian, X., Sharma, V., Zhuang, H., Kalinin, S. V....

    2020

  30. [38]

    Wang, H.-C., Botti, S., and Marques, M. A. L. (January, 2021) Predicting stable crystalline compounds using chemical similarity. npj Computational Materials, 7(1), 1–9 35 Publisher: Nature Publishing Group

    2021

  31. [39]

    (October, 2016) Computational Screening of All Stoichiometric Inorganic Materials

    Davies, D., Butler, K., Jackson, A., Morris, A., Frost, J., Skelton, J., and Walsh, A. (October, 2016) Computational Screening of All Stoichiometric Inorganic Materials. Chem, 1(4), 617–627

    2016

  32. [40]

    Matbench Discovery: An evaluation framework for machine learning crys- tal stability prediction

    Riebesell, J., Goodall, R., Benner, P., Chiang, Y., Deng, B., Lee, A., Jain, A., and Persson, K. Matbench Discovery: An evaluation framework for machine learning crys- tal stability prediction. (August, 2023) Available at: https://janosh.github.io/ matbench-discovery

    2023

  33. [41]

    (April, 2023) Available at: https://figshare.com/ articles/dataset/Matbench_Discovery_v1_0_0/22715158/13

    Matbench Discovery v1.0.0. (April, 2023) Available at: https://figshare.com/ articles/dataset/Matbench_Discovery_v1_0_0/22715158/13. 36 Supplemental Material Synthesizability score distribution for all iterations (a) (b) Fig. 1: Synthesizability score distribution for Iteratio...

    2023 arXiv

  34. [2018]

    The Journal of Physical Chemistry B, 122(2), 625–632 Publisher: American Chemical Society

    Materials Screening for the Discovery of New Half-Heuslers: Machine Learning 32 versus ab Initio Methods. The Journal of Physical Chemistry B, 122(2), 625–632 Publisher: American Chemical Society

  35. [2023]

    npj Computational Materials, 9(1), 1–9 Number: 1 Publisher: Nature Publishing Group

    A critical examination of robustness and generalizability of machine learning prediction of materials properties. npj Computational Materials, 9(1), 1–9 Number: 1 Publisher: Nature Publishing Group. 33

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.