Pith. sign in

REVIEW 3 major objections 5 minor 31 references

Multi-task parallelism for robust pre-training of graph foundation models on multi-source, multi-fidelity atomistic modeling data

T0 review · 3 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 Multi-task learning with per-dataset output heads makes graph foundation models for atomistic data more accurate and transferable, and a new head-level parallelism scheme lets them scale to 24 million structures.

desk verdict A real scalability idea and the largest MTL pre-training run to date, but Table 2's identical rows and the undocumented energy alignment keep the force-accuracy claim from being trusted as printed. read the letter →

arxiv 2506.21788 v1 pith:4NTLLZYM submitted 2025-06-26 cs.LG cond-mat.mtrl-scics.AIphysics.atm-clus

Massimiliano Lupo Pasini , Jong Youl Choi , Pei Zhang , Kshitij Mehta , Rylie Weaver , Ashwin M. Aji , Karl W. Schulz , Jorda Polo
show 1 more author
Prasanna Balaprakash
This is my paper · ORCID
classification cs.LGcond-mat.mtrl-scics.AIphysics.atm-clus
keywords graphneuralnetworksmulti-tasklearningmodelparallelismdistributeddatamulti-fidelityatomisticmodelingfoundationmodelssupercomputerscaling
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

Single-dataset pre-training fails to generalize across organic and inorganic chemistry, so a foundation model for atomistic data needs to learn from many sources at once. This paper argues that multi-task learning—one shared message-passing encoder with a separate output head for each dataset—is the right way to do that: pre-training on 24 million structures from five datasets with different DFT settings yields lower energy and force errors, and better transferability, than either dataset-specific models or one model trained on all data mixed together. The paper also introduces multi-task parallelism, which places each dataset head on its own GPU and synchronizes only the shared encoder, so memory per GPU no longer grows with the number of datasets. They report near-ideal strong scaling on up to 320 GPUs and weak scaling on up to 1,920 GPUs across three US supercomputers, suggesting that adding datasets to a foundation model can be as simple as adding GPUs for their heads.

What carries the argument

The central object is the two-level hierarchical multi-task architecture in the authors' HydraGNN graph neural network. The first level of multi-task learning splits a shared message-passing encoder into one branch per dataset; the second level splits each branch into a head for energy per atom and a head for atomic forces. The load-bearing mechanism is multi-task parallelism: each process holds a copy of the shared encoder plus one complete dataset head, heads run forward and backward passes concurrently, and only the encoder gradients are averaged across processes during synchronization. This changes the memory footprint per GPU from $P_s + N_h P_h$ to $P_s + P_h$, so the number of datasets becomes an axis that scales independently of model depth and width. Process sub-groups carry out distributed data parallelism inside each head while a global group synchronizes the shared layers.

What would settle it

Re-run GFM-MTL-All with the energy-per-atom alignment replaced by arbitrary per-dataset constants (or by offsets computed from a different reference total energy), and compare out-of-distribution MAE on the five test sets against Tables 1 and 2; if the MAE values stay essentially unchanged, the MTL advantage is an artifact of the alignment, while if they degrade sharply, the alignment is doing real work.

Watch

Extended reading notes

Core claim

The central claim is that multi-task learning should be the default pre-training scheme for graph foundation models on multi-source, multi-fidelity atomistic data, because it simultaneously gives high in-distribution accuracy and out-of-distribution transferability where single-dataset training and single-head all-data training fail. The authors report that a two-level MTL model—dataset-specific branches, each split into energy and force heads—achieves the best or near-best error on every one of the five datasets considered, for example cutting energy error on the inorganic MPTrj set from 0.4248 for the all-data baseline to 0.0627. They further claim that multi-task parallelism, integrated with distributed data parallelism through process sub-groups, makes this method scalable: memory per GPU drops from $P_s + N_h P_h$ to $P_s + P_h$, strong scaling stays near ideal up to 320 GPUs for large effective batches, and weak scaling holds up to 1,920 GPUs on the three supercomputers tested.

Load-bearing premise

The energy-per-atom values across the five datasets were 'consistently aligned,' but the paper never describes how the alignment offsets were obtained; if those offsets are not physically correct, the per-dataset heads could simply absorb dataset-specific biases, making the reported transferability gains an artifact of the alignment rather than of multi-task learning.

Editorial extensions

If this is right

  • A single pre-trained GFM can serve both organic and inorganic downstream tasks, because the shared encoder learns common chemistry while each head specializes in its source dataset's data distribution.
  • Adding a new dataset to pre-training costs one additional GPU (or group of GPUs) for its head rather than a larger model or more memory per GPU, so the approach directly addresses the flood of new atomistic datasets.
  • The simultaneous prediction of energy per atom and atomic forces from the same shared representation means downstream users do not need separate models for the two quantities.
  • If the memory-scaling argument holds, the method becomes more attractive as the number of heads grows relative to the shared encoder, which is the regime typical of message-passing GNNs with many datasets.
  • Strong and weak scaling results imply that pre-training on hundreds of millions of structures—the paper's stated next step—is computationally reachable with current exascale systems.

Reading between the lines

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

  • The reported accuracy gains depend on the unstated energy-per-atom alignment across the five datasets; a control experiment with randomized or permuted offsets would separate the effect of multi-task learning from the effect of that alignment.
  • Multi-task parallelism is not specific to graph neural networks: any multi-task architecture with independent per-task heads could use the same head-splitting scheme, so the method may transfer to other scientific foundation models.
  • The scaling comparison is against the same method without head-level parallelism; a comparison with tensor or pipeline parallelism under the same memory constraints would clarify when each strategy should be chosen.
  • A direct stress test is the paper's own future-work target of 359 million structures covering all natural elements: if the accuracy and scaling results persist at that size, the main practical obstacle to a universal atomistic foundation model is removed.
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 model-parallelism strategy for multi-task learning (MTL) in graph foundation models for atomistic simulation. The method, implemented in HydraGNN, partitions dataset-specific output heads across GPUs while keeping the shared encoder synchronized via distributed data parallelism. The authors pre-train on five datasets totaling over 24 million structures (ANI1x, QM7-X, Transition1x, MPTrj, Alexandria), report energy and force MAE for seven model variants (Tables 1 and 2), and present weak- and strong-scaling experiments on Perlmutter, Aurora, and Frontier. The central claims are that MTL improves accuracy and transferability relative to single-dataset training and that the proposed 2D parallelization scales efficiently.

Significance. If the numerical results hold, the paper makes a useful engineering contribution: it demonstrates a practical way to scale MTL-based graph foundation model pre-training to many heterogeneous datasets, and it provides open-source artifacts and benchmarks across three major supercomputing platforms. The aggregation of 24 million structures from diverse organic and inorganic datasets is itself a notable resource. However, the accuracy and transferability claims rest on Tables 1 and 2, which contain an internal inconsistency in the force-prediction results, and on an energy-alignment procedure that is not described. These issues must be resolved before the scientific claims can be accepted. The scaling results are plausible and interesting, but they are not the main source of the paper's claimed scientific advance.

major comments (3)
  1. [Table 2] The GFM-Baseline-All row in Table 2 is identical, to four decimal places, to the Model-MPTrj row (0.0508, 1.1734, 0.1597, 0.0039, 0.1425), and the Model-Alexandria row differs only in the MPTrj column (0.0038 vs. 0.0039). A single model trained on all five datasets cannot produce exactly the same force MAE on all five test sets as a model trained only on MPTrj. This indicates a likely copy/paste error or evaluation with the wrong checkpoint. Because the paper's force-accuracy and transferability claims rely directly on this table, the force-prediction comparisons are invalidated until the table is corrected and the evaluation reproduced.
  2. [Sec. 4.1] The text states that 'we consistently aligned the energy per atom values across all the datasets' but provides no description of how the offsets were determined. If the per-dataset offsets are fitted to minimize training error, the MTL heads can absorb dataset-specific energy biases, which would confound the comparison between GFM-MTL-All and the single-dataset baselines in Table 1. The authors should specify the alignment method, state whether it uses only training data, and ideally show the sensitivity of the downstream MAE results to the alignment procedure.
  3. [Sec. 5.1, Tables 1 and 2] All MAE numbers in Tables 1 and 2 are reported without error bars or repeated-run statistics. Several differences that support the central claim are small, such as the MPTrj energy MAE of 0.0627 for GFM-MTL-All versus 0.0651 for Model-MPTrj, or the Alexandria force MAE of 0.0039 for GFM-MTL-All versus 0.0038 for Model-Alexandria. Without variance estimates or significance tests, the claim that MTL 'significantly improves' accuracy is not statistically supported. The authors should report mean and standard deviation over multiple seeds, or at least provide a clear justification for why single runs are sufficient.
minor comments (5)
  1. [Sec. 5] The name 'NERSC-Pelrmutter' contains a typo and should be 'Perlmutter'.
  2. [Sec. 6] The conclusion contains 'NERSC=Perlmutter' and 'using using'; these should be fixed.
  3. [Sec. 4.3] The memory-scaling expressions 'Ps + (Ph * Nh)' and '(Ps + Ph)' assume that the number of processes equals the number of heads; this should be stated explicitly, and the notation Ps, Ph, Nh should be defined in one place.
  4. [Fig. 4] The scaling plots would benefit from a legend distinguishing MTL-base and MTL-par in all subplots, and from a description of how the plotted time is averaged over the three reported epochs.
  5. [Sec. 3] The phrase 'denominated as' is used where 'denoted as' is intended; this occurs several times in Sections 5.1 and 5.2.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the empirical MTL comparison is self-contained; the Table 2 duplication is a correctness issue, not a circular step.

full rationale

The paper makes no formal derivation claims: the central results are empirical MAE values from trained models evaluated on held-out test partitions. The claimed advantage of two-level MTL is measured against single-dataset models and a mixed-data baseline, providing independent grounding rather than reducing to a fitted parameter or a self-citation. Self-citations to earlier HydraGNN work ([12]–[14]) supply the architecture and an earlier hyperparameter search, but those are configuration inputs and are not used to define the accuracy numbers. The energy-per-atom alignment mentioned in Sec. 4.1 ('We consistently aligned the energy per atom values across all the datasets') is not described, which is a missing-support concern, not a self-referential one. The identical force-MAE rows for GFM-Baseline-All, Model-MPTrj, and Model-Alexandria in Table 2 are an internal inconsistency in the reported evidence that should be investigated, but an empirical reporting error is not a circular derivation. No equation, definition, or self-citation chain makes any 'prediction' equivalent to its input by construction. Therefore the circularity score is 0.

Assumptions & free parameters 2 free parameters · 3 assumptions · 0 invented entities

The central claim depends on a handful of domain assumptions and at least one unreported preprocessing parameter (energy alignment). No new physical entities are introduced.

free parameters (2)
  • Per-dataset energy per atom offset = unknown
    Section 4.1 states energies were consistently aligned across datasets, but the offsets or alignment method are not reported. If estimated from data, these offsets are free parameters that could absorb dataset biases.
  • Model hyperparameters (hidden units, head units, learning rate, local batch size) = 866 / 889 / 0.001 / 128
    Chosen from an earlier hyperparameter search on different data, not re-tuned here; these choices affect the results but are not justified for the new dataset.
assumptions (3)
  • domain assumption DFT and CCSD energies and forces in the five datasets are accurate enough to serve as ground truth
    The pre-training quality depends on the physical validity of the source calculations (Sec. 4.1).
  • domain assumption A single shared EGNN architecture can represent energy and force surfaces across organic and inorganic compounds
    The entire approach assumes transferable representations across chemical space (Sec. 3, Sec. 5.1).
  • domain assumption The five datasets, after alignment, provide consistent labels for joint multi-task training
    MTL assumes per-dataset heads can share an encoder; the alignment is asserted but not demonstrated (Sec. 4.1).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multi-task parallelism for robust pre-training of graph foundation models on multi-source, multi-fidelity atomistic modeling data." pith.science (2026). https://pith.science/paper/4NTLLZYM

@misc{pith2026250621788,
  author       = {Pith},
  title        = {Pith review of: Multi-task parallelism for robust pre-training of graph foundation models on multi-source, multi-fidelity atomistic modeling data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/4NTLLZYM}},
  note         = {Machine review of arXiv:2506.21788}
}
read the original abstract

Graph foundation models using graph neural networks promise sustainable, efficient atomistic modeling. To tackle challenges of processing multi-source, multi-fidelity data during pre-training, recent studies employ multi-task learning, in which shared message passing layers initially process input atomistic structures regardless of source, then route them to multiple decoding heads that predict data-specific outputs. This approach stabilizes pre-training and enhances a model's transferability to unexplored chemical regions. Preliminary results on approximately four million structures are encouraging, yet questions remain about generalizability to larger, more diverse datasets and scalability on supercomputers. We propose a multi-task parallelism method that distributes each head across computing resources with GPU acceleration. Implemented in the open-source HydraGNN architecture, our method was trained on over 24 million structures from five datasets and tested on the Perlmutter, Aurora, and Frontier supercomputers, demonstrating efficient scaling on all three highly heterogeneous super-computing architectures.

Figures

Figures reproduced from arXiv: 2506.21788 by the authors.

Figure 1
Figure 1. Heatmap that describes the frequency of occurrence of each element of the periodic table across data sampled resulting from the aggregation of the datasets ANI1x, QM7-X, MPTrj, Alexandria, and Transition1x [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. HydraGNN architecture that implements a two-level MTL approach. The first level of MTL splits the architecture into multiple output decoders to process data coming from different sources, whereas the second level of MTL further splits each output decoder into two sub-decoders for the simultaneous prediction of energies and atomic forces. heads. Therefore, forward propagations for each MTL head can be performed concu… view at source ↗
Figure 3
Figure 3. Left: Partitioning of one single HydraGNN model replica across N pro￾cesses. Each process is mapped to one NVIDIA A100 GPU on NERSC-Perlmutter, one AMD Mi250X GCD on OLCF-Frontier, and one Intel Data Center GPU Tile on ALCF-Aurora. Right: Organization of processes into sub-process groups using torch.DeviceMesh. One global group is used to perform DDP to synchronize the gra￾dients associated with the MPNN encoder par… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Scaling experiment results. The plots show weak scaling (left) and strong scal￾ing (right) on Frontier (top), Perlmutter (middle), and Aurora (bottom). The y-axis shows the average total training time per epoch, including data loading, forward, and backward passes, whi…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 15 canonical work pages

  1. [1]

    npj Computational Materials10, 154 (2024)

    Allen, A.E.A., Lubbers, N., Matin, S., Smith, J., Messerly, R., Tretiak, S., Barros, K.:Learningtogether:Towardsfoundationmodelsformachinelearninginteratomic potentials with meta-learning. npj Computational Materials10, 154 (2024)

    work page 2024

  2. [2]

    Barroso-Luque, L., Shuaibi, M., Fu, X., Wood, B.M., Dzamba, M., Gao, M., Rizvi, A., Zitnick, C.L., Ulissi, Z.W.: Open materials 2024 (omat24) inorganic materials dataset and models (2024), https://arxiv.org/abs/2410.12771

    arXiv 2024

  3. [3]

    In: Proceedings of the Twelfth International Conference on Learning Representations (2024), https://openreview.net/forum?id=Zc2aIcucwc

    Beaini, D., Huang, S., Cunha, J.A., Li, Z., Moisescu-Pareja, G., Dymov, O., Maddrell-Mander, S., McLean, C., Wenkel, F., Müller, L., Mohamud, J.H., Parviz, A., Craig, M., Koziarski, M., Lu, J., Zhu, Z., Gabellini, C., Klaser, K., Dean, J., Wognum, C., Sypetkowski, M., Rabusseau, G., Rabbany, R., Tang, J., Mor- ris, C., Ravanelli, M., Wolf, G., Tossou, P.,...

    work page 2024

  4. [4]

    Advanced Func- tional Materials 34, 2404043 (2024)

    Cerqueira, T.F.T., Fang, Y.W., Errea, I., Sanna, A., Marques, M.A.L.: Searching materials space for hydride superconductors at ambient pressure. Advanced Func- tional Materials 34, 2404043 (2024). https://doi.org/10.1002/adfm.202404043

    doi:10.1002/adfm.202404043 2024

  5. [5]

    Oak Ridge National Laboratory (ORNL), Oak Ridge, TN (United States) (11 2023)

    Choi, J.Y., Lupo Pasini, M., Zhang, P., Mehta, K., Liu, F., Bae, J., Ibrahim, K.: Ddstore: Distributed data store for scalable training of graph neural networks on large atomistic modeling datasets. Oak Ridge National Laboratory (ORNL), Oak Ridge, TN (United States) (11 2023). https://doi.org/10.1145/3624062.3624171, https://www.osti.gov/biblio/2251635

    arXiv 2023

  6. [6]

    a framework for high-performance data manage- ment

    Godoy, W., Podhorszki, N., Wang, R., Atkins, C., Eisenhauer, G., Gu, J., Davis, P., Choi, J.Y., Germaschewski, K., Huck, K., et al.: Adios 2: The adapt- able input output system. a framework for high-performance data manage- ment. SoftwareX 12(1) (07 2020). https://doi.org/10.1016/j.softx.2020.100561, https://www.osti.gov/biblio/1638101

    work page arXiv 2020

  7. [7]

    Journal of Machine Learning Research13(25), 723–773 (2012), https://jmlr.csail.mit.edu/papers/v13/gretton12a.html

    Gretton, A., Borgwardt, K.M., Rasch, M.J., Schölkopf, B., Smola, A.J.: A kernel two-sample test. Journal of Machine Learning Research13(25), 723–773 (2012), https://jmlr.csail.mit.edu/papers/v13/gretton12a.html

    work page 2012

  8. [8]

    Scientific Data 8, 43 (2021)

    Hoja, J., Sandonas, L.M., Ernst, B.G., Vazquez-Mayagoitia, A., DiStasio Jr., R.A., Tkatchenko, A.: QM7-X, a comprehensive dataset of quantum-mechanical properties spanning the chemical space of small organic molecules. Scientific Data 8, 43 (2021). https://doi.org/https://doi.org/10.1038/s41597-021-00812-2, https://www.nature.com/articles/s41597-021-00812-2

    doi:10.1038/s41597-021-00812-2 2021

Show all 31 references
  1. [9]

    ChemRxiv (2023)

    Jacobson, L., Stevenson, J., Ramezanghorbani, F., Dajnowicz, S., Leswing, K.: Leveraging multitask learning to improve the transferability of machine learned force fields. ChemRxiv (2023). https://doi.org/10.26434/chemrxiv-2023-8n737, preprint. This content has not been peer-reviewed

    2023 doi

  2. [10]

    Lupo Pasini et al

    Jain, A., Ong, S.P., Hautier, G., Chen, W., Richards, W.D., Dacek, S., Cholia, S., Gunter, D., Skinner, D., Ceder, G., Persson, K.A.: Commentary: The Ma- 14 M. Lupo Pasini et al. terials Project: A materials genome approach to accelerating materials innova- tion. APL Materials...

    2013 doi

  3. [11]

    In: Proceedings of the SC ’23 Workshops of The Inter- national Conference on High Performance Computing, Network, Storage, and Analysis

    Lee, K.L.K., Gonzales, C., Spellings, M., Galkin, M., Miret, S., Kumar, N.: Towards foundation models for materials science: The open matsci ml toolkit. In: Proceedings of the SC ’23 Workshops of The Inter- national Conference on High Performance Computing, Network, Storage, a...

    2023

  4. [12]

    Journal of Supercomputing81, Article 618 (Mar 2025), open Access; Published: 14 March 2025

    Lupo Pasini, M., Choi, J.Y., Mehta, K., Zhang, P., Rogers, D., Bae, J., Ibrahim, K.Z., Aji, A.M., Schulz, K.W., Polo, J., Balaprakash, P.: Scalable training of trust- worthy and energy-efficient predictive graph foundation models for atomistic ma- terials modeling: a case stud...

    2025

  5. [13]

    https://doi.org/10.2172/2224153, https://www.osti.gov/biblio/2224153

    Lupo Pasini, M., Choi, J.Y., Zhang, P., Baker, J.: User Manual - HydraGNN: Distributed PyTorch Implementation of Multi-Headed Graph Convolutional Neural Networks (11 2023). https://doi.org/10.2172/2224153, https://www.osti.gov/biblio/2224153

    2023

  6. [14]

    https://doi.org/10.11578/dc.20240131.1, https://www.osti.gov/biblio/2283293

    Lupo Pasini, M., Choi, J.Y., Zhang, P., Baker, J.: Hydragnn v3.0, version v3.0 (2 2024). https://doi.org/10.11578/dc.20240131.1, https://www.osti.gov/biblio/2283293

    2024

  7. [15]

    In: Proceedings of the International Conference for High Performance Computing, Network- ing, Storage and Analysis

    Md, V., Misra, S., Ma, G., Mohanty, R., Georganas, E., Heinecke, A., Kalamkar, D., Ahmed, N.K., Avancha, S.: Distgnn: scalable dis- tributed training for large-scale graph neural networks. In: Proceedings of the International Conference for High Performance Computing, Network-...

    2021

  8. [16]

    Advanced Materials35, 2210788 (mar 2023)

    Schmidt, J., Hoffmann, N., Wang, H.C., Borlido, P., Carriço, P.J.M.A., Cerqueira, T.F.T., Botti, S., Marques, M.A.L.: Machine-learning-assisted determination of the global zero-temperature phase diagram of materials. Advanced Materials35, 2210788 (mar 2023). https://doi.org/10...

    2023 doi

  9. [17]

    Sci- ence Advances 7(49), eabi7948 (2021)

    Schmidt, J., Pettersson, L., Verdozzi, C., Botti, S., Marques, M.A.L.: Crys- tal graph attention networks for the prediction of stable materials. Sci- ence Advances 7(49), eabi7948 (2021). https://doi.org/10.1126/sciadv.abi7948, https://www.science.org/doi/abs/10.1126/sciadv.abi7948

    2021 doi

  10. [18]

    Scientific Data 9, 64 (2022)

    Schmidt, J., Wang, H.C., Cerqueira, T.F.T., Botti, S., Marques, M.A.L.: A dataset of 175k stable and metastable materials calculated with the PBEsol and SCAN functionals. Scientific Data 9, 64 (2022). https://doi.org/0.1038/s41597- 022-01177-w

    2022

  11. [19]

    Scientific Data 9, 779 (2022)

    Schreiner, M., Bhowmik, A., Vegge, T., Busk, J., Winther, O.: Transition1x - a dataset for building generalizable reactive machine learning potentials. Scientific Data 9, 779 (2022)

    2022

  12. [20]

    Shiota, T., Ishihara, K., Do, T.M., Mori, T., Mizukami, W.: Taming multi-domain, -fidelity data: Towards foundation models for atomistic scale simulations (2024), https://arxiv.org/abs/2412.13088

    2024

  13. [21]

    In: The Twelfth International Conference on Learning Repre- sentations (2024), https://openreview.net/forum?id=PfPnugdxup

    Shoghi, N., Kolluru, A., Kitchin, J.R., Ulissi, Z.W., Zitnick, C.L., Wood, B.M.: From molecules to materials: Pre-training large generalizable models for atomic Title Suppressed Due to Excessive Length 15 property prediction. In: The Twelfth International Conference on Learnin...

    2024

  14. [22]

    Nature Com- munications 10(1), 2903 (2019)

    Smith, J.S., Nebgen, B.T., Zubatyuk, R., Lubbers, N., Devereux, C., Barros, K., Tretiak, S., Isayev, O., Roitberg, A.E.: Approaching coupled cluster accuracy with a general-purpose neural network potential through transfer learning. Nature Com- munications 10(1), 2903 (2019). ...

    2019 doi

  15. [23]

    Scientific Data 7, 134 (2020)

    Smith, J.S., Zubatyuk, R., Nebgen, B., Lubbers, N., Barros, K., Roit- berg, A.E., Isayev, O., Tretiak, S.: The ANI-1ccx and ANI-1x data sets, coupled-cluster and density functional theory properties for molecules. Scientific Data 7, 134 (2020). https://doi.org/https://doi.org/...

    2020 doi

  16. [24]

    In: Proceedings of the European Conference on Computer Vision (ECCV) Work- shops

    Sun,B.,Saenko,K.:Deepcoral:Correlationalignmentfordeepdomainadaptation. In: Proceedings of the European Conference on Computer Vision (ECCV) Work- shops. pp. 443–450. Springer (2016). https://doi.org/10.1007/978-3-319-49409- 8_35

    2016 doi

  17. [25]

    In: The Thirty- eighth Annual Conference on Neural Information Processing Systems (2024), https://openreview.net/forum?id=klqhrq7fvB

    Sypetkowski, M., Wenkel, F., Poursafaei, F., Dickson, N., Suri, K., Fradkin, P., Beaini, D.: On the scalability of GNNs for molecular graphs. In: The Thirty- eighth Annual Conference on Neural Information Processing Systems (2024), https://openreview.net/forum?id=klqhrq7fvB

    2024

  18. [26]

    In: Proceedings of the AI for Accelerated Materials Design Workshop at NeurIPS 2023 (2023), https://openreview.net/forum?id=EiT2bLsfM9

    Takeda, S., Priyadarsini, I., Kishimoto, A., Shinohara, H., Hamada, L., Hirose, M., Fuchiwaki, J., Nakano, D.: Multi-modal foundation model for material design. In: Proceedings of the AI for Accelerated Materials Design Workshop at NeurIPS 2023 (2023), https://openreview.net/f...

    2023

  19. [27]

    https://doi.org/10.1088/2053-1583/accc43

    Wang, H.C., Schmidt, J., Marques, M.A.L., Wirtz, L., Romero, A.H.: Symmetry- basedcomputationalsearchfornovelbinaryandternary2dmaterials.2DMaterials 10(3), 035007 (2023). https://doi.org/10.1088/2053-1583/accc43

    2023 doi

  20. [28]

    In: Proceedings of the Sixteenth European Conference on Com- puter Systems

    Wang, L., Yin, Q., Tian, C., Yang, J., Chen, R., Yu, W., Yao, Z., Zhou, J.: Flexgraph: a flexible and efficient distributed framework for gnn training. In: Proceedings of the Sixteenth European Conference on Com- puter Systems. p. 67–82. EuroSys ’21, Association for Computing ...

    2021

  21. [29]

    In: Pro- ceedings of the 25th ACM SIGKDD International Conference on Knowledge Dis- covery & Data Mining

    Yang, H.: Aligraph: A comprehensive graph neural network platform. In: Pro- ceedings of the 25th ACM SIGKDD International Conference on Knowledge Dis- covery & Data Mining. p. 3165–3166. KDD ’19, Association for Computing Ma- chinery, New York, NY, USA (2019). https://doi.org/...

    2019

  22. [30]

    Zhang, D., Liu, X., Zhang, X., Zhang, C., Cai, C., Bi, H., Du, Y., Qin, X., Peng, A., Huang, J., Li, B., Shan, Y., Zeng, J., Zhang, Y., Liu, S., Li, Y., Chang, J., Wang, X., Zhou, S., Liu, J., Luo, X., Wang, Z., Jiang, W., Wu, J., Yang, Y., Yang, J., Yang, M., Gong, F.Q., Zhan...

    2024 arXiv

  23. [31]

    In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discov- ery and Data Mining

    Zheng, D., Song, X., Yang, C., LaSalle, D., Karypis, G.: Distributed hybrid cpu and gpu training for graph neural networks on billion-scale heterogeneous graphs. In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discov- ery and Data Mining. p. 4582–4591. KDD ’22, ...

    2022

Pith tools

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