Pith. sign in

REVIEW 4 major objections 7 minor 33 references

LaMM: Semi-Supervised Pre-Training of Large-Scale Materials Models

T0 review · 4 major / 7 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read Semi-supervised pretraining on ~300M samples improves materials models' fine-tuning speed and accuracy.

desk verdict Solid engineering with a credible load-balancing result, but the semi-supervised claim is untested because the unlabeled subset is never ablated. read the letter →

arxiv 2505.22208 v1 pith:443B5PQS submitted 2025-05-28 cs.LG

Yosuke Oyama , Yusuke Majima , Eiji Ohta , Yasufumi Sakai This is my paper · ORCID
classification cs.LG
keywords neuralnetworkpotentialssemi-supervisedpre-trainingdenoisingself-supervisedlearningloadbalancingmaterialsinformaticsfoundationmodelcoordinate
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

LaMM is a semi-supervised pre-training method for neural network potentials (NNPs) that pools publicly available DFT-labeled datasets with 94 million unlabeled PubChem3D conformers into a joint training set of about 300 million samples. The paper claims that this joint pretraining produces a single model which, after fine-tuning, reaches better energy and force accuracy on the HME21 benchmark than training from scratch, while requiring about 2.5x fewer fine-tuning steps. To enable the mix, it introduces a universal loss over per-dataset heads that accepts any combination of energy, force, and unlabeled samples, plus a denoising task whose label is the mean-subtracted coordinate displacement. It also contributes a load-balancing algorithm that sorts samples by atom count and restructures mini-batches so heterogeneous datasets train efficiently on GPUs.

What carries the argument

The argument is carried by three components: (1) a universal head/loss design in which the model predicts energy and force for every dataset with masked losses, so labeled and unlabeled samples share one encoder; (2) a mean-subtracted denoising task for unlabeled structures that assigns unique per-atom labels and avoids translational ambiguity; (3) a load-balancing shuffler that sorts samples by atom count, chunks them by GPU count, and greedily assigns them to batches, eliminating idle GPUs and out-of-memory stalls.

What would settle it

Train the same model on the joint dataset with the PubChem3D subset removed, then fine-tune on HME21; if the energy and force MAE plus the speedups are indistinguishable from the full LaMM pretraining, the unlabeled denoising data contributes nothing and the claimed semi-supervised benefit collapses.

Watch

Extended reading notes

Core claim

The central claim is that pre-training a single NNP on a ~300M semi-labeled dataset that includes unlabeled PubChem3D conformers improves fine-tuning speed and accuracy on downstream datasets relative to training from scratch. On HME21, LaMM-S (PaiNN) reaches 21.1 meV/atom energy MAE versus 23.7 for the scratch PaiNN, and LaMM-L (EquiformerV2) reaches 20.2 versus 24.4, while both cut the fine-tuning steps needed to reach threshold by about 2.5x for energy and 1.9x-2.6x for forces. The paper argues the gains come from the breadth of structures and partial labels in the joint dataset, not from any single subset.

Load-bearing premise

The method's advantage depends on the unstated assumption that predicting mean-subtracted coordinate displacements on force-field conformers teaches a representation that transfers to true DFT energies and forces under a different functional.

Editorial extensions

If this is right

  • Pre-trained LaMM checkpoints can be fine-tuned on new DFT functionals with less compute than training from scratch, lowering the cost of bespoke potentials.
  • The universal loss means future datasets with partial labels (energy-only, force-only, or unlabeled) can be folded into pretraining without relabeling.
  • The mean-subtracted denoising trick is a drop-in improvement for any coordinate-denoising self-supervised pretraining, not just this dataset mix.
  • The load-balancing algorithm generalizes to other graph neural network workloads with high variance in system size, enabling larger effective batch sizes on memory-constrained models.

Reading between the lines

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

  • Because the paper compares joint pretraining only against scratch, the specific contribution of the unlabeled PubChem3D subset is not isolated; an ablation with PubChem3D removed would clarify whether the gains come from unlabeled data or from the larger label diversity.
  • The denoising task is tied to MMFF94s force-field geometries, so its transfer to DFT targets may depend on how close those geometries are to relaxed DFT structures; testing on conformers from a different generator would delimit its range.
  • If the load-balancing algorithm makes training throughput scale with dataset size rather than worst-case atom counts, then pretraining on much larger unlabeled collections becomes computationally feasible, an avenue the paper implies but does not test.
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

4 major / 7 minor

Summary. LaMM proposes a semi-supervised pre-training framework for neural network potentials. It combines ~282M samples from ten open datasets, including 94M unlabeled PubChem3D conformers, into one joint training set, using a multi-head loss that handles heterogeneous labels (energy, forces, or none) and a mean-subtracted coordinate denoising pseudo-label. Two models, LaMM-S (PaiNN-based) and LaMM-L (EquiformerV2-based), are pre-trained and then fine-tuned on the HME21 dataset. The paper reports that the pre-trained models improve fine-tuning accuracy and speed over scratch training, and also presents a load-balancing algorithm that reduces step-time variance and GPU load imbalance.

Significance. If the semi-supervised benefit were cleanly isolated, LaMM would be a valuable contribution: it shows how unlabeled structural databases can be folded into NNP pre-training, and the load-balancing trick addresses a real engineering bottleneck for large graph neural networks. The mean-subtraction in the denoising task is simple, compelling, and supported by Fig. 7 as an internal convergence improvement. The paper also demonstrates that pre-training on a large heterogeneous joint dataset transfers to a previously unseen functional (PBE) and structure type. However, the strongest claim—that the unlabeled PubChem3D data and the denoising task are responsible for the downstream gains—is not tested. The evaluation is confined to one downstream dataset, and the speedup numbers rest on self-chosen thresholds. These gaps prevent the paper from supporting its headline claims as written.

major comments (4)
  1. [§4.4, Table 3] The central 'semi-supervised' claim is not isolated. The only comparison is pre-training on the full joint dataset (labeled subsets plus PubChem3D with denoising loss) versus training from scratch. Because the labeled subsets already contain about 188M S2EF-Total samples from OC20/OC22/ODAC23 plus Transition1x, QM7-X, QMugs and IS2RE variants, and because JMP-style supervised pre-training on OC20 is known to transfer (Shoghi et al., 2024), the observed HME21 improvements could be entirely due to supervised pre-training. An ablation that pre-trains on the labeled subsets alone—and, ideally, also with the baseline (non-mean-subtracted) denoising loss—is necessary to attribute the gains to the unlabeled data and to the proposed denoising label.
  2. [§4.4, Table 3 and Abstract] Only one downstream benchmark (HME21) is reported, yet the abstract claims 'improved fine-tuning performance in terms of both speed and accuracy' and the introduction promises performance after fine-tuning 'for various downstream datasets.' HME21 is a single dataset of disordered multi-element systems computed with PBE; it does not substantiate a claim about general transferability. Either add results on additional downstream tasks (e.g., OC20 IS2RE/S2EF, QM9, MD17, or a materials relaxation benchmark) or substantially temper the generalization claim.
  3. [§4.2, Fig. 7] The experimental evidence for the proposed denoising labeling is internal to the self-supervised task: Fig. 7 shows that the mean-subtracted denoising objective converges faster on the PubChem3D validation error. This does not demonstrate that the resulting encoder representation improves downstream energy/force prediction. The paper should show that checkpoints trained with the proposed denoising loss fine-tune to better HME21 (or other) accuracy than checkpoints trained with the baseline denoising loss, or otherwise restrict the claim to 'improved SSL convergence.'
  4. [§4.4, Table 3] The reported fine-tuning speedups depend on self-selected thresholds (25 meV/atom and 210 meV/Å for LaMM-S; 25 meV/atom and 150 meV/Å for LaMM-L) and no steps-to-threshold or wall-clock times are given. Without the underlying learning curves or a table listing the number of steps each model needs to reach each threshold, the speedup factors (2.51x, 1.89x, 2.50x, 2.61x) are not verifiable and may be sensitive to threshold choice. Please report the full fine-tuning curves or the explicit step counts for each model and threshold.
minor comments (7)
  1. [Abstract and Table 2] The abstract says '~300 million semi-labeled samples' but Table 2 lists 282M samples (432M with temperature sampling). Use consistent numbers and explain what 'semi-labeled' means.
  2. [Table 2] The 'Total' row lists '79' under the average-atoms column with no label; clarify that this is the weighted average number of atoms, or split the row into columns that are individually explained.
  3. [§4.1] The throughput improvements of 2.44x and 3.38x on 4 nodes appear to include both the load-balancing algorithm and an increased batch size (B'=16 or B'=8 versus B'=4 or B'=2). Separate the effect of load balancing from that of the larger batch size, or state explicitly that the quoted figures are the combined gain.
  4. [§4.2, Fig. 7] The y-axis label 'Noise norm. MAE [ ]' has an empty unit; since atom coordinates are in Ångström, the unit should be Å.
  5. [§3.2, Eq. (5)] The force loss is described as 'system size-balanced L2 MAE losses,' but Eq. (5) uses the L2 norm of the force difference, which is not an MAE. Clarify the terminology or write the correct formula.
  6. [§4.3] LaMM-L is pre-trained for only 0.16 epochs. A more completely pre-trained checkpoint would strengthen the comparison, and the 'experimentally trained' wording suggests the current checkpoint may be provisional.
  7. [§5] The conclusion calls LaMM an 'omnipotent foundation NNP model.' This overstates the empirical scope, which is a single downstream task and a small set of element combinations; temper the language.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity; the semi-supervised attribution gap is an experimental validity concern, not a circular derivation.

full rationale

I walked the derivation chain. The claimed benefit is evaluated by fine-tuning pre-trained checkpoints (0.1/2 epoch LaMM-S, 0.16 epoch LaMM-L) versus scratch on HME21, an external dataset with PBE labels not included in pre-training data; the accuracy result (Table 3) is therefore not a restatement of a training loss or a fitted quantity. The denoising-label improvement (Fig. 7) is validated against held-out PubChem3D denoising error, which is an internal SSL metric but not the paper's final downstream claim; it supports only the auxiliary task. The load-balancing claim is measured directly by step-time distributions on the same hardware. I found no equation that reduces to its input, no fitted parameter renamed as a prediction, and no load-bearing self-citation chain: references to JMP, OC20/OC22, and Feng et al. are external prior work. The main evidentiary weakness is that no ablation removes PubChem3D or the denoising loss, so the specific contribution of the unlabeled subset to HME21 gains is not isolated; that is an experimental attribution gap, not circularity under the stated definitions.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claims rest on a set of training hyperparameters and design choices (loss weights, temperature, split count, thresholds) whose values are not always reported or justified. The denoising pseudo-label concept is borrowed from prior work, and its transfer to real force prediction is a domain assumption. No new physical entities are introduced, and the pseudo-force labels are not new physical objects.

free parameters (5)
  • Loss weights lambda_E and lambda_F = not stated
    Weights in Eq. (5) for energy and force losses; values are not reported in the paper, presumably inherited from JMP [24].
  • Temperature sampling temperature T = 2
    Chosen to balance subset sizes in the joint dataset (Section 3.1); a hand-picked hyperparameter affecting training distribution.
  • Number of split subsets S in load balancing = 10000
    The first step of the load-balancing algorithm splits each dataset into 10,000 subsets (Section 4.1); value chosen without stated sensitivity analysis.
  • Fine-tuning speedup thresholds = 25 meV/atom, 210 meV/angstrom (LaMM-S); 25 meV/atom, 150 meV/angstrom (LaMM-L)
    Thresholds in Table 3 used to define speedup; no justification for these particular values, making the speedup metric threshold-dependent.
  • Pre-training epochs = 2 (LaMM-S), 0.16 (LaMM-L)
    The models are stopped early for resource reasons; the 0.1-epoch LaMM-S checkpoint is also used in Table 3. The choice of checkpoint is arbitrary.
assumptions (4)
  • domain assumption Total energy is the sum of per-atom energies (additivity).
    Assumed in Eqs. (1)-(3) as the standard NNP architecture; not validated for the materials in the joint dataset.
  • domain assumption Denoising pseudo-forces (negative random displacements) provide a learning signal correlated with the true force field.
    Motivated by prior work [25] and stated in Section 3.3; this transfer is the core justification for using unlabeled PubChem3D data.
  • ad hoc to paper Temperature sampling (T=2) creates a training distribution from which downstream tasks such as HME21 are well represented.
    The choice of T=2 and the strategy of fixing the largest subset and repeating others are design choices (Section 3.1) not derived from any optimality criterion.
  • domain assumption The load-balancing algorithm does not change model outputs or the loss function, only the order of samples.
    Stated in Section 3.4 and used to argue that system performance improvements do not affect model accuracy.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LaMM: Semi-Supervised Pre-Training of Large-Scale Materials Models." pith.science (2026). https://pith.science/paper/443B5PQS

@misc{pith2026250522208,
  author       = {Pith},
  title        = {Pith review of: LaMM: Semi-Supervised Pre-Training of Large-Scale Materials Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/443B5PQS}},
  note         = {Machine review of arXiv:2505.22208}
}
abstract

Neural network potentials (NNPs) are crucial for accelerating computational materials science by surrogating density functional theory (DFT) calculations. Improving their accuracy is possible through pre-training and fine-tuning, where an NNP model is first pre-trained on a large-scale dataset and then fine-tuned on a smaller target dataset. However, this approach is computationally expensive, mainly due to the cost of DFT-based dataset labeling and load imbalances during large-scale pre-training. To address this, we propose LaMM, a semi-supervised pre-training method incorporating improved denoising self-supervised learning and a load-balancing algorithm for efficient multi-node training. We demonstrate that our approach effectively leverages a large-scale dataset of $\sim$300 million semi-labeled samples to train a single NNP model, resulting in improved fine-tuning performance in terms of both speed and accuracy.

Figures

Figures reproduced from arXiv: 2505.22208 by the authors.

Figure 1
Figure 1. Overview of LaMM. “E”: total energy, “E′”: adsorption energy after relaxation, “F”: forces, “N”: coordinate noises. The blue parts are our proposed methods explained in this paper. 6 [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Heatmap of the number of datasets containing each element. Markers [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Existing and proposed denoising labeling algorithm. [PITH_FULL_IMAGE:figures/full_fig_p011_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Histogram of the per-sample number of atoms of our pre-training [PITH_FULL_IMAGE:figures/full_fig_p012_4.png]
Figure 5
Figure 5. Figure 5: Our proposed load balancing algorithm. Red/blue samples contain [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Training step time with and without our load-balancing method on [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: Learning curves for self-supervised training on the PubChem3D [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: UMAP [32] node-level feature map of the pre-training dataset. PaiNN’s rotation-invariant representations with a length of 512 dimensions are used. 1000 atoms are randomly sampled from each subset. Top left: colored by subset. Top right: colored by element with colors d…
Figure 9
Figure 9. Figure 9: Learning curves of LaMM-S for HME21 fine-tuning. 31-points mov [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 31 canonical work pages

  1. [1]

    Equivariant message passing for the prediction of tensorial properties and molecular spectra

    Kristof Sch¨ utt, Oliver Unke, and Michael Gastegger. Equivariant message passing for the prediction of tensorial properties and molecular spectra. In International Conference on Machine Learning , pages 9377–9388. PMLR, 2021

    work page 2021

  2. [2]

    Chgnet as a pretrained uni- versal neural network potential for charge-informed atomistic modelling

    Bowen Deng, Peichen Zhong, KyuJung Jun, Janosh Riebesell, Kevin Han, Christopher J Bartel, and Gerbrand Ceder. Chgnet as a pretrained uni- versal neural network potential for charge-informed atomistic modelling. Nature Machine Intelligence , 5(9):1031–1041, 2023

    work page 2023

  3. [3]

    Equiformerv2: Improved equivariant transformer for scaling to higher- degree representations

    Yi-Lun Liao, Brandon M Wood, Abhishek Das, and Tess Smidt. Equiformerv2: Improved equivariant transformer for scaling to higher- degree representations. In The Twelfth International Conference on Learn- ing Representations, 2024

    work page 2024

  4. [4]

    A universal graph deep learning interatomic potential for the periodic table

    Chi Chen and Shyue Ping Ong. A universal graph deep learning interatomic potential for the periodic table. Nature Computational Science, 2(11):718– 728, 2022

    work page 2022

  5. [5]

    Mat- tersim: A deep learning atomistic model across elements, temperatures and pressures, 2024

    Han Yang, Chenxi Hu, Yichi Zhou, Xixian Liu, Yu Shi, Jielan Li, Guanzhi Li, Zekun Chen, Shuizhou Chen, Claudio Zeni, Matthew Horton, Robert Pinsler, Andrew Fowler, Daniel Z¨ ugner, Tian Xie, Jake Smith, Lixin Sun, Qian Wang, Lingyu Kong, Chang Liu, Hongxia Hao, and Ziheng Lu. Mat- tersim: A deep learning atomistic model across elements, temperatures and p...

    work page 2024

  6. [6]

    E (3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials

    Simon Batzner, Albert Musaelian, Lixin Sun, Mario Geiger, Jonathan P Mailoa, Mordechai Kornbluth, Nicola Molinari, Tess E Smidt, and Boris Kozinsky. E (3)-equivariant graph neural networks for data-efficient and accurate interatomic potentials. Nature communications, 13(1):2453, 2022

    work page 2022

  7. [7]

    Teanet: Universal neural network interatomic potential inspired by iterative electronic relaxations

    So Takamoto, Satoshi Izumi, and Ju Li. Teanet: Universal neural network interatomic potential inspired by iterative electronic relaxations. Compu- tational Materials Science , 207:111280, 2022

    work page 2022

  8. [8]

    Schnet: A continuous-filter convolutional neural network for modeling quantum inter- actions

    Kristof Sch¨ utt, Pieter-Jan Kindermans, Huziel Enoc Sauceda Felix, Ste- fan Chmiela, Alexandre Tkatchenko, and Klaus-Robert M¨ uller. Schnet: A continuous-filter convolutional neural network for modeling quantum inter- actions. In I. Guyon, U. Von Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, editors, Advances in Neural Info...

    work page 2017

Show all 33 references
  1. [9]

    Do transformers really perform badly for graph representation? Advances in neural information processing systems, 34:28877–28888, 2021

    Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, and Tie-Yan Liu. Do transformers really perform badly for graph representation? Advances in neural information processing systems, 34:28877–28888, 2021

    2021

  2. [10]

    The open catalyst challenge 2021: Compe- tition report

    Abhishek Das, Muhammed Shuaibi, Aini Palizhati, Siddharth Goyal, Aditya Grover, Adeesh Kolluru, Janice Lan, Ammar Rizvi, Anuroop Sri- ram, Brandon M Wood, et al. The open catalyst challenge 2021: Compe- tition report. In NeurIPS (Competition and Demos) , pages 29–40, 2021

    2021

  3. [11]

    Wood, Misko Dzamba, Meng Gao, Ammar Rizvi, C

    Luis Barroso-Luque, Muhammed Shuaibi, Xiang Fu, Brandon M. Wood, Misko Dzamba, Meng Gao, Ammar Rizvi, C. Lawrence Zitnick, and Zachary W. Ulissi. Open materials 2024 (omat24) inorganic materials dataset and models, 2024

    2024

  4. [12]

    Towards universal neural network potential for mate- rial discovery applicable to arbitrary combination of 45 elements

    So Takamoto, Chikashi Shinagawa, Daisuke Motoki, Kosuke Nakago, Wen- wen Li, Iori Kurata, Taku Watanabe, Yoshihiro Yayama, Hiroki Iriguchi, Yusuke Asano, et al. Towards universal neural network potential for mate- rial discovery applicable to arbitrary combination of 45 elemen...

    2022

  5. [13]

    End-to-end symmetry preserving inter-atomic potential energy model for finite and extended systems

    Linfeng Zhang, Jiequn Han, Han Wang, Wissam Saidi, Roberto Car, et al. End-to-end symmetry preserving inter-atomic potential energy model for finite and extended systems. Advances in neural information processing systems, 31, 2018

    2018

  6. [14]

    Directional message passing for molecular graphs

    Johannes Gasteiger, Janek Groß, and Stephan G¨ unnemann. Directional message passing for molecular graphs. In International Conference on Learning Representations (ICLR), 2020

    2020

  7. [15]

    Gemnet: universal directional graph neural networks for molecules

    Johannes Klicpera, Florian Becker, and Stephan G¨ unnemann. Gemnet: universal directional graph neural networks for molecules. In Proceedings of the 35th International Conference on Neural Information Processing Sys- tems, NIPS ’21, Red Hook, NY, USA, 2024. Curran Associates Inc. 21

    2024

  8. [16]

    Lawrence Zitnick, and Abhishek Das

    Johannes Gasteiger, Muhammed Shuaibi, Anuroop Sriram, Stephan G¨ unnemann, Zachary Ward Ulissi, C. Lawrence Zitnick, and Abhishek Das. Gemnet-OC: Developing graph neural networks for large and di- verse molecular simulation datasets. Transactions on Machine Learning Research, 2022

    2022

  9. [17]

    Lawrence Zitnick, and Zachary Ulissi

    Lowik Chanussot, Abhishek Das, Siddharth Goyal, Thibaut Lavril, Muhammed Shuaibi, Morgane Riviere, Kevin Tran, Javier Heras-Domingo, Caleb Ho, Weihua Hu, Aini Palizhati, Anuroop Sriram, Brandon Wood, Junwoong Yoon, Devi Parikh, C. Lawrence Zitnick, and Zachary Ulissi. Open cat...

    2020

  10. [18]

    Wood, Sid- dharth Goyal, Abhishek Das, Javier Heras-Domingo, Adeesh Kolluru, Am- mar Rizvi, Nima Shoghi, Anuroop Sriram, F´ elix Therrien, Jehad Abed, Oleksandr Voznyy, Edward H

    Richard Tran, Janice Lan, Muhammed Shuaibi, Brandon M. Wood, Sid- dharth Goyal, Abhishek Das, Javier Heras-Domingo, Adeesh Kolluru, Am- mar Rizvi, Nima Shoghi, Anuroop Sriram, F´ elix Therrien, Jehad Abed, Oleksandr Voznyy, Edward H. Sargent, Zachary Ulissi, and C. Lawrence Zi...

    2022

  11. [19]

    Brabson, Abhishek Das, Zachary Ulissi, Matt Uyttendaele, Andrew J

    Anuroop Sriram, Sihoon Choi, Xiaohan Yu, Logan M. Brabson, Abhishek Das, Zachary Ulissi, Matt Uyttendaele, Andrew J. Medford, and David S. Sholl. The open dac 2023 dataset and challenges for sorbent discovery in direct air capture. ACS Central Science, 10(5):923–941, 2024

    2023

  12. [20]

    Elena, D´ avid P

    Ilyes Batatia, Philipp Benner, Yuan Chiang, Alin M. Elena, D´ avid P. Kov´ acs, Janosh Riebesell, Xavier R. Advincula, Mark Asta, Matthew Avaylon, William J. Baldwin, Fabian Berger, Noam Bernstein, Arghya Bhowmik, Samuel M. Blau, Vlad C˘ arare, James P. Darby, Sandip De, Flavi...

    2024

  13. [21]

    Mace: Higher order equivariant message passing neural networks 22 for fast and accurate force fields

    Ilyes Batatia, David P Kovacs, Gregor Simm, Christoph Ortner, and Gabor Csanyi. Mace: Higher order equivariant message passing neural networks 22 for fast and accurate force fields. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neur...

    2022

  14. [22]

    York, Shi Liu, Tong Zhu, Zhicheng Zhong, Jian Lv, Jun Cheng, Weile Jia, Mohan Chen, Guolin Ke, Weinan E, Linfeng Zhang, and Han Wang

    Duo Zhang, Xinzijian Liu, Xiangyu Zhang, Chengqian Zhang, Chun Cai, Hangrui Bi, Yiming Du, Xuejian Qin, Anyang Peng, Jiameng Huang, Bowen Li, Yifan Shan, Jinzhe Zeng, Yuzhi Zhang, Siyuan Liu, Yifan Li, Junhan Chang, Xinyan Wang, Shuo Zhou, Jianchuan Liu, Xiaoshan Luo, Zhenyu W...

    2024

  15. [23]

    https://www.preferred.jp/en/news/p r20240917/

    PFN and ENEOS Release v7 of PFP Neural Network Potential for Univer- sal Atomistic Simulator Matlantis. https://www.preferred.jp/en/news/p r20240917/

  16. [24]

    Kitchin, Zachary Ward Ulissi, C

    Nima Shoghi, Adeesh Kolluru, John R. Kitchin, Zachary Ward Ulissi, C. Lawrence Zitnick, and Brandon M Wood. From molecules to materials: Pre-training large generalizable models for atomic property prediction. In The Twelfth International Conference on Learning Representations , 2024

    2024

  17. [25]

    Fractional denoising for 3d molecular pre-training

    Shikun Feng, Yuyan Ni, Yanyan Lan, Zhi-Ming Ma, and Wei-Ying Ma. Fractional denoising for 3d molecular pre-training. In International Con- ference on Machine Learning , pages 9938–9961. PMLR, 2023

    2023

  18. [26]

    Pub- chem3d: a new resource for scientists

    Evan E Bolton, Jie Chen, Sunghwan Kim, Lianyi Han, Siqian He, Wenyao Shi, Vahan Simonyan, Yan Sun, Paul A Thiessen, Jiyao Wang, et al. Pub- chem3d: a new resource for scientists. Journal of cheminformatics , 3:1–15, 2011

    2011

  19. [27]

    Transition1x-a dataset for building generalizable reactive machine learning potentials

    Mathias Schreiner, Arghya Bhowmik, Tejs Vegge, Jonas Busk, and Ole Winther. Transition1x-a dataset for building generalizable reactive machine learning potentials. Scientific Data , 9(1):779, 2022

    2022

  20. [28]

    Qm7-x, a comprehensive dataset of quantum-mechanical properties span- ning the chemical space of small organic molecules

    Johannes Hoja, Leonardo Medrano Sandonas, Brian G Ernst, Alvaro Vazquez-Mayagoitia, Robert A DiStasio Jr, and Alexandre Tkatchenko. Qm7-x, a comprehensive dataset of quantum-mechanical properties span- ning the chemical space of small organic molecules. Scientific data, 8(1):43, 2021

    2021

  21. [29]

    Qmugs, quantum mechanical properties of drug-like molecules

    Clemens Isert, Kenneth Atz, Jos´ e Jim´ enez-Luna, and Gisbert Schneider. Qmugs, quantum mechanical properties of drug-like molecules. Scientific Data, 9(1):273, 2022

    2022

  22. [30]

    https://github.com/FAIR-Chem/fairchem/p ull/267

    Load balancing #267, 2021. https://github.com/FAIR-Chem/fairchem/p ull/267. 23

    2021

  23. [31]

    Abci 2.0: Advances in open ai computing infrastructure at aist

    Shinichiro Takizawa, Yusuke Tanimura, Hidemoto Nakada, Ryousei Takano, and Hirotaka Ogawa. Abci 2.0: Advances in open ai computing infrastructure at aist. IPSJ SIG, techreport HPC-180 , 2021

    2021

  24. [32]

    Umap: Uniform manifold approximation and projection

    Leland McInnes, John Healy, Nathaniel Saul, and Lukas Grossberger. Umap: Uniform manifold approximation and projection. The Journal of Open Source Software, 3(29):861, 2018

    2018

  25. [33]

    http: //www.jmol.org/

    Jmol: an open-source Java viewer for chemical structures in 3D. http: //www.jmol.org/. 24

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.