pith. sign in

arxiv: 2605.20581 · v1 · pith:ZI4IHMLRnew · submitted 2026-05-20 · 💻 cs.LG · cond-mat.mtrl-sci

TriForces: Augmenting Atomistic GNNs for Transferable Representations

Ali Ramlaoui , Alexandre Duval , Hannah Bull , Victor Schmidt , Hugues Talbot , Fragkiskos D. Malliaros , Joseph Musielewicz This is my paper

Pith reviewed 2026-05-21 07:07 UTC · model grok-4.3

classification 💻 cs.LG cond-mat.mtrl-sci
keywords atomistic GNNstransferable representationsself-supervised learningmachine learning interatomic potentialscomposition and structure separationlimited data regimematerials benchmarksMLIPs
0
0 comments X

The pith

TriForces separates composition and structure streams in atomistic GNNs with self-supervised learning to improve transferable representations across chemical domains.

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

The paper aims to show that machine learning interatomic potentials transfer inconsistently across domains because their representations often lose accessible composition and structure information. TriForces introduces a model-agnostic three-stream framework that processes composition and structure separately and adds self-supervised objectives to keep this information preserved. A sympathetic reader would care because MLIPs must frequently adapt to new target chemistries using only small and expensive task-specific datasets, and better transfer would reduce reliance on large DFT computations. The reported results include gains on MatBench and QM9 without DFT labels, efficient structure retrieval in the latent space, and a 57 percent reduction in energy mean absolute error at 20,000 samples on OMat24.

Core claim

TriForces is a model-agnostic three-stream framework that separates composition and structure information, combined with self-supervised learning to preserve transferable representations. This leads to improved performance on MatBench and QM9 over baselines without needing DFT labels, enables efficient similar structure retrieval through its learned latent space, and on OMat24 in the limited-data regime reduces energy MAE by 57 percent at 20K samples while improving force MAE across sample sizes.

What carries the argument

The three-stream architecture that explicitly separates composition and structure processing streams alongside self-supervised learning objectives to maintain accessible information in the representations.

If this is right

  • Models achieve better accuracy on materials property prediction tasks using limited training data from new domains.
  • Pretrained models transfer more effectively to target chemistries without requiring additional large-scale DFT calculations.
  • The learned latent space supports efficient retrieval of similar atomic structures.
  • The framework applies directly to multiple existing MLIP architectures with released pretrained variants.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Similar separation of data streams could generalize to other scientific machine learning settings that involve multimodal inputs such as molecular graphs and sequences.
  • Improved latent representations might lower the sample complexity needed for active learning loops in materials discovery workflows.
  • The approach could support more efficient screening of large chemical spaces by enabling better zero-shot or few-shot adaptation to unseen element combinations.

Load-bearing premise

The assumption that explicitly separating composition and structure streams plus self-supervised objectives will preserve and improve transferable information across domains, rather than the gains coming from increased model capacity or from the particular choice of auxiliary tasks.

What would settle it

Training variants with and without the composition-structure separation and self-supervised components on the OMat24 limited-data regime, then checking whether the reported 57 percent energy MAE reduction at 20K samples disappears, would test whether the framework itself drives the transferable gains.

read the original abstract

Machine learning interatomic potentials (MLIPs) achieve excellent accuracy when trained on large Density Functional Theory (DFT) data. To be useful in practice, they must often be adapted to target chemistries using small and expensive task-specific datasets. However, MLIPs transfer inconsistently across domains, with representations that often loose accessible composition and structure information. To address this, we present TriForces, a model-agnostic three-stream framework that separates composition and structure information, combined with self-supervised learning to preserve transferable representations. TriForces improves performance on MatBench and QM9 over baselines without needing DFT labels and enables efficient similar structure retrieval through its learned latent space. On OMat24, in limited-data training regime, TriForces reduces energy MAE by 57% at 20K samples only and improves force MAE across sample sizes. We release pretrained TriForces variants across multiple MLIP architectures with code at https://github.com/Ramlaoui/triforces.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper introduces TriForces, a model-agnostic three-stream framework for atomistic GNNs that explicitly separates composition and structure information streams and augments them with self-supervised learning objectives to preserve transferable representations. It claims concrete improvements including a 57% reduction in energy MAE at 20K samples on OMat24 in the limited-data regime, better force MAE across sample sizes, gains on MatBench and QM9 without DFT labels, and utility for similar structure retrieval via the learned latent space. Pretrained variants and code are released.

Significance. If the performance gains are shown to arise specifically from the composition/structure separation and SSL objectives rather than capacity increases or auxiliary task choice, the work could meaningfully advance transfer learning for ML interatomic potentials, enabling more efficient adaptation to new chemistries with small datasets and reducing dependence on large DFT corpora. The open release of models and code supports reproducibility and follow-on work.

major comments (2)
  1. [Abstract / Experiments] Abstract and Experiments section: The central claim of improved transferable representations via the three-stream design plus SSL rests on the 57% energy MAE reduction at 20K samples on OMat24 and cross-benchmark gains, yet no parameter counts, capacity-matched single-stream baselines, or ablations isolating the streams from the auxiliary objectives are reported. This leaves open the possibility that gains derive from increased total capacity or particular SSL task selection rather than the proposed separation.
  2. [Experiments] §4 (or equivalent experimental setup): No details are provided on hyperparameter matching across baselines, statistical significance of the reported MAE improvements, or whether the three-stream architecture was compared against equivalently parameterized single-stream models with the same SSL losses. These controls are load-bearing for attributing the limited-data regime gains to the architectural innovation.
minor comments (2)
  1. [Abstract] The abstract states that representations 'often loose accessible composition and structure information' but does not quantify this loss or show how the three-stream design recovers it beyond the downstream metrics.
  2. [Results] Figure or table presenting the OMat24 results should include error bars or multiple runs to support the cross-sample-size force MAE improvements.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive comments. We address each major comment below and commit to revisions that strengthen the evidence for our claims.

read point-by-point responses

  1. Referee: [Abstract / Experiments] Abstract and Experiments section: The central claim of improved transferable representations via the three-stream design plus SSL rests on the 57% energy MAE reduction at 20K samples on OMat24 and cross-benchmark gains, yet no parameter counts, capacity-matched single-stream baselines, or ablations isolating the streams from the auxiliary objectives are reported. This leaves open the possibility that gains derive from increased total capacity or particular SSL task selection rather than the proposed separation.

    Authors: We agree that explicit parameter counts and capacity-matched controls are necessary to attribute gains specifically to the three-stream separation and SSL objectives. In the revised manuscript we will add a table listing total parameter counts for TriForces and all baselines. We will also include new ablation experiments that compare the full three-stream model against single-stream architectures with matched parameter budgets, both with and without the auxiliary SSL losses, on the OMat24 limited-data splits. revision: yes

  2. Referee: [Experiments] §4 (or equivalent experimental setup): No details are provided on hyperparameter matching across baselines, statistical significance of the reported MAE improvements, or whether the three-stream architecture was compared against equivalently parameterized single-stream models with the same SSL losses. These controls are load-bearing for attributing the limited-data regime gains to the architectural innovation.

    Authors: We will expand the experimental details section to document the hyperparameter search and matching protocol used for all models. We will also report mean and standard deviation of energy and force MAEs over multiple independent runs. In addition, we will add direct comparisons of the three-stream model against single-stream baselines that employ identical SSL losses and equivalent total parameter counts, allowing clearer isolation of the architectural contribution. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical performance claims rest on held-out benchmarks

full rationale

The paper presents TriForces as a three-stream augmentation to existing GNN architectures, separating composition and structure streams plus auxiliary self-supervised losses, then reports measured MAE reductions on OMat24 (held-out limited-data splits), MatBench, and QM9. No equations, uniqueness theorems, or derivation steps are shown that reduce by construction to the inputs; the performance numbers are external test-set statistics rather than re-statements of the model definition or fitted parameters. The architecture choices are presented as design decisions with empirical validation, not as forced by self-citation or ansatz smuggling. The result is therefore self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

The central claim rests on the untested premise that explicit stream separation plus self-supervised objectives will yield better transfer than standard GNN training; no free parameters, axioms, or invented entities are visible in the abstract.

pith-pipeline@v0.9.0 · 5722 in / 1296 out tokens · 33823 ms · 2026-05-21T07:07:45.692075+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

89 extracted references · 89 canonical work pages

  1. [1]

    2020 , journal =

    An Introduction to Electrocatalyst Design using Machine Learning for Renewable Energy Storage , author =. 2020 , journal =

    work page 2020

  2. [2]

    EquiformerV2: Improved Equivariant Transformer for Scaling to Higher-Degree Representations , booktitle =

    Yi. EquiformerV2: Improved Equivariant Transformer for Scaling to Higher-Degree Representations , booktitle =. 2024 , url =

    work page 2024

  3. [3]

    2025 , journal =

    Orb-v3: atomistic simulation at scale , author =. 2025 , journal =

    work page 2025

  4. [4]

    Ilyes Batatia and Dávid Péter Kovács and Gregor N. C. Simm and Christoph Ortner and Gábor Csányi , year =

    work page

  5. [5]

    arXiv preprint arXiv:2508.20875 , year=

    LeMat-Traj: A Scalable and Unified Dataset of Materials Trajectories for Atomistic Modeling , author=. arXiv preprint arXiv:2508.20875 , year=

    work page arXiv

  6. [6]

    2023 , journal =

    CHGNet: Pretrained universal neural network potential for charge-informed atomistic modeling , author =. 2023 , journal =

    work page 2023

  7. [7]

    2024 , journal =

    Open Materials 2024 (OMat24) Inorganic Materials Dataset and Models , author =. 2024 , journal =

    work page 2024

  8. [8]

    2024 , journal =

    Generalizing Denoising to Non-Equilibrium Structures Improves Equivariant Force Fields , author =. 2024 , journal =

    work page 2024

  9. [9]

    2025 , journal =

    A Foundational Potential Energy Surface Dataset for Materials , author =. 2025 , journal =

    work page 2025

  10. [10]

    2023 , journal =

    Matbench Discovery - A framework to evaluate machine learning crystal stability predictions , author =. 2023 , journal =

    work page 2023

  11. [11]

    APL materials , volume=

    Commentary: The Materials Project: A materials genome approach to accelerating materials innovation , author=. APL materials , volume=. 2013 , publisher=

    work page 2013

  12. [12]

    Handbook of Materials Modeling: Methods: Theory and Modeling , pages=

    The materials project: Accelerating materials design through theory-driven data and tools , author=. Handbook of Materials Modeling: Methods: Theory and Modeling , pages=. 2020 , publisher=

    work page 2020

  13. [13]

    Jom , volume=

    Materials design and discovery with high-throughput density functional theory: the open quantum materials database (OQMD) , author=. Jom , volume=. 2013 , publisher=

    work page 2013

  14. [14]

    Jonathan Schmidt and Love Pettersson and Claudio Verdozzi and Silvana Botti and Miguel A. L. Marques , title =. Science Advances , volume =. 2021 , doi =. https://www.science.org/doi/pdf/10.1126/sciadv.abi7948 , abstract =

    work page doi:10.1126/sciadv.abi7948 2021

  15. [15]

    Materials Today Physics , volume=

    Improving machine-learning models in materials science through large datasets , author=. Materials Today Physics , volume=. 2024 , publisher=

    work page 2024

  16. [16]

    2022 , journal =

    Tutorial on amortized optimization , author =. 2022 , journal =

    work page 2022

  17. [17]

    Computational Materials Science , volume=

    The AFLOW library of crystallographic prototypes: Part 4 , author=. Computational Materials Science , volume=. 2024 , publisher=

    work page 2024

  18. [18]

    2023 , journal =

    From Molecules to Materials: Pre-training Large Generalizable Models for Atomic Property Prediction , author =. 2023 , journal =

    work page 2023

  19. [19]

    Scientific data , volume=

    OPTIMADE, an API for exchanging materials data , author=. Scientific data , volume=. 2021 , publisher=

    work page 2021

  20. [20]

    Jäger and Eiaki V

    Lauri Himanen and Marc O.J. Jäger and Eiaki V. Morooka and Filippo. DScribe: Library of descriptors for machine learning in materials science , journal =. 2020 , issn =. doi:https://doi.org/10.1016/j.cpc.2019.106949 , url =

    work page doi:10.1016/j.cpc.2019.106949 2020

  21. [21]

    AI for Accelerated Materials Design - ICLR 2025 , year=

    LeMat-Bulk: aggregating, and de-duplicating quantum chemistry materials databases , author=. AI for Accelerated Materials Design - ICLR 2025 , year=

    work page 2025

  22. [22]

    arXiv preprint arXiv:2505.01590 , year=

    Multi-fidelity learning for interatomic potentials: Low-level forces and high-level energies are all you need , author=. arXiv preprint arXiv:2505.01590 , year=

    work page arXiv

  23. [23]

    Journal of Chemical Theory and Computation , volume=

    The good, the bad, and the ugly: pseudopotential inconsistency errors in molecular applications of density functional theory , author=. Journal of Chemical Theory and Computation , volume=. 2023 , publisher=

    work page 2023

  24. [24]

    Proceedings of the 2021 conference on empirical methods in natural language processing: system demonstrations , pages=

    Datasets: A community library for natural language processing , author=. Proceedings of the 2021 conference on empirical methods in natural language processing: system demonstrations , pages=

    work page 2021

  25. [25]

    2018 , issn =

    Matminer: An open source toolkit for materials data mining , journal =. 2018 , issn =. doi:https://doi.org/10.1016/j.commatsci.2018.05.018 , url =

    work page doi:10.1016/j.commatsci.2018.05.018 2018

  26. [26]

    Draxl and M

    Draxl, Claudia and Scheffler, Matthias , title =. 2019 , month =. doi:10.1088/2515-7639/ab13bb , url =

    work page doi:10.1088/2515-7639/ab13bb 2019

  27. [27]

    Journal of Physics: Condensed Matter , volume=

    Ask Hjorth Larsen and Jens Jørgen Mortensen and Jakob Blomqvist and Ivano E Castelli and Rune Christensen and Marcin Dułak and Jesper Friis and Michael N Groves and Bjørk Hammer and Cory Hargus and Eric D Hermes and Paul C Jennings and Peter Bjerre Jensen and James Kermode and John R Kitchin and Esben Leonhard Kolsbjerg and Joseph Kubal and Kristen Kaasbj...

    work page 2017

  28. [28]

    A hitchhiker’s guide to geometric gnns for 3d atomic systems.arXiv preprint arXiv:2312.07511, 2023

    A hitchhiker's guide to geometric gnns for 3d atomic systems , author=. arXiv preprint arXiv:2312.07511 , year=

    work page arXiv

  29. [29]

    International Conference on Machine Learning , pages=

    Faenet: Frame averaging equivariant gnn for materials modeling , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    work page 2023

  30. [30]

    Computational Materials Science , volume=

    Python Materials Genomics (pymatgen): A robust, open-source python library for materials analysis , author=. Computational Materials Science , volume=. 2013 , publisher=

    work page 2013

  31. [31]

    The Journal of Chemical Physics , volume =

    Batatia, Ilyes and Benner, Philipp and Chiang, Yuan and others , title =. The Journal of Chemical Physics , volume =. 2025 , month =. doi:10.1063/5.0297006 , url =

    work page doi:10.1063/5.0297006 2025

  32. [32]

    Nature communications , volume=

    SpookyNet: Learning force fields with electronic degrees of freedom and nonlocal effects , author=. Nature communications , volume=. 2021 , publisher=

    work page 2021

  33. [33]

    2017 , journal =

    SchNet - a deep learning architecture for molecules and materials , author =. 2017 , journal =

    work page 2017

  34. [34]

    Scientific data , volume=

    The ANI-1ccx and ANI-1x data sets, coupled-cluster and density functional theory properties for molecules , author=. Scientific data , volume=. 2020 , publisher=

    work page 2020

  35. [35]

    Forces are not Enough: Benchmark and Critical Evaluation for Machine Learning Force Fields with Molecular Simulations , author =. Trans. Mach. Learn. Res. , year =

    work page

  36. [36]

    Chanussot and Abhishek Das and Siddharth Goyal and Thibaut Lavril and Muhammed Shuaibi and M

    L. Chanussot and Abhishek Das and Siddharth Goyal and Thibaut Lavril and Muhammed Shuaibi and M. Rivière and Kevin Tran and Javier Heras-Domingo and Caleb Ho and Weihua Hu and Aini Palizhati and Anuroop Sriram and Brandon Wood and Junwoong Yoon and Devi Parikh and C. L. Zitnick and Zachary W. Ulissi , journal =. The Open Catalyst 2020 (

    work page 2020

  37. [37]

    2022 , doi =

    Pymatviz: visualization toolkit for materials informatics , author =. 2022 , doi =

    work page 2022

  38. [38]

    1808.01590 , archivePrefix=

    Atsushi Togo and Kohei Shinohara and Isao Tanaka , year=. 1808.01590 , archivePrefix=

    work page arXiv

  39. [39]

    2025 , journal =

    Energy & Force Regression on DFT Trajectories is Not Enough for Universal Machine Learning Interatomic Potentials , author =. 2025 , journal =

    work page 2025

  40. [40]

    2024 , journal =

    Orb: A Fast, Scalable Neural Network Potential , author =. 2024 , journal =

    work page 2024

  41. [41]

    Chemical Reviews , volume=

    Machine learning force fields , author=. Chemical Reviews , volume=. 2021 , publisher=

    work page 2021

  42. [42]

    Chemical Reviews , volume=

    Data generation for machine learning interatomic potentials and beyond , author=. Chemical Reviews , volume=. 2024 , publisher=

    work page 2024

  43. [43]

    2025 , journal =

    The Open Molecules 2025 (OMol25) Dataset, Evaluations, and Models , author =. 2025 , journal =

    work page 2025

  44. [44]

    Kitchin and Daniel S

    Brandon M Wood and Misko Dzamba and Xiang Fu and Meng Gao and Muhammed Shuaibi and Luis Barroso-Luque and Kareem Abdelmaqsoud and Vahe Gharakhanyan and John R. Kitchin and Daniel S. Levine and Kyle Michel and Anuroop Sriram and Taco Cohen and Abhishek Das and Sushree Jagriti Sahoo and Ammar Rizvi and Zachary Ward Ulissi and C. Lawrence Zitnick , booktitle...

    work page 2025

  45. [45]

    npj Computational Materials , volume=

    Accelerating materials discovery using artificial intelligence, high performance computing and robotics , author=. npj Computational Materials , volume=. 2022 , publisher=

    work page 2022

  46. [46]

    npj Computational Materials , volume=

    Training data selection for accuracy and transferability of interatomic potentials , author=. npj Computational Materials , volume=. 2022 , publisher=

    work page 2022

  47. [47]

    The Journal of Chemical Physics , volume=

    Challenges for machine learning force fields in reproducing potential energy surfaces of flexible molecules , author=. The Journal of Chemical Physics , volume=. 2021 , publisher=

    work page 2021

  48. [48]

    npj Computational Materials , volume=

    Universal machine learning interatomic potentials are ready for phonons , author=. npj Computational Materials , volume=. 2025 , publisher=

    work page 2025

  49. [49]

    Nature Materials , volume=

    Machine-learned potentials for next-generation matter simulations , author=. Nature Materials , volume=. 2021 , publisher=

    work page 2021

  50. [50]

    AI for Accelerated Materials Design-NeurIPS 2025 , year=

    Superior Molecular Representations from Intermediate Encoder Layers , author=. AI for Accelerated Materials Design-NeurIPS 2025 , year=

    work page 2025

  51. [51]

    Journal of Chemical Theory and Computation , year=

    Transferability of Data Sets between Machine-Learned Interatomic Potential Algorithms , author=. Journal of Chemical Theory and Computation , year=

    work page

  52. [52]

    npj Computational Materials , volume=

    Cross-functional transferability in foundation machine learning interatomic potentials , author=. npj Computational Materials , volume=. 2025 , publisher=

    work page 2025

  53. [53]

    ACS Applied Materials & Interfaces , volume=

    Performance assessment of universal machine learning interatomic potentials: Challenges and directions for materials’ surfaces , author=. ACS Applied Materials & Interfaces , volume=. 2024 , publisher=

    work page 2024

  54. [54]

    Nature , volume=

    A generative model for inorganic materials design , author=. Nature , volume=. 2025 , publisher=

    work page 2025

  55. [55]

    2020 , journal =

    Fast and Uncertainty-Aware Directional Message Passing for Non-Equilibrium Molecules , author =. 2020 , journal =

    work page 2020

  56. [56]

    Proceedings of the 42nd International Conference on Machine Learning , pages =

    Learning Smooth and Expressive Interatomic Potentials for Physical Property Prediction , author =. Proceedings of the 42nd International Conference on Machine Learning , pages =. 2025 , editor =

    work page 2025

  57. [57]

    Generalizing Denoising to Non-Equilibrium Structures Improves Equivariant Force Fields , journal =

    Yi. Generalizing Denoising to Non-Equilibrium Structures Improves Equivariant Force Fields , journal =. 2024 , url =

    work page 2024

  58. [58]

    2025 , journal =

    Understanding and Enhancing Mask-Based Pretraining towards Universal Representations , author =. 2025 , journal =

    work page 2025

  59. [59]

    2022 , journal =

    Graph Contrastive Learning for Materials , author =. 2022 , journal =

    work page 2022

  60. [60]

    International Conference on Learning Representations , year=

    Simple GNN Regularisation for 3D Molecular Property Prediction and Beyond , author=. International Conference on Learning Representations , year=

    work page

  61. [61]

    Nature communications , volume=

    Predicting materials properties without crystal structure: deep representation learning from stoichiometry , author=. Nature communications , volume=. 2020 , publisher=

    work page 2020

  62. [62]

    Npj Computational Materials , volume=

    Compositionally restricted attention-based network for materials property predictions , author=. Npj Computational Materials , volume=. 2021 , publisher=

    work page 2021

  63. [63]

    npj Computational Materials , volume=

    Crystal twins: self-supervised learning for crystalline material property prediction , author=. npj Computational Materials , volume=. 2022 , publisher=

    work page 2022

  64. [64]

    Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    Emerging properties in self-supervised vision transformers , author=. Proceedings of the IEEE/CVF International Conference on Computer Vision , pages=

    work page

  65. [65]

    International Conference on Machine Learning , pages=

    A simple framework for contrastive learning of visual representations , author=. International Conference on Machine Learning , pages=. 2020 , organization=

    work page 2020

  66. [66]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Momentum contrast for unsupervised visual representation learning , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    work page

  67. [67]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Masked autoencoders are scalable vision learners , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    work page

  68. [68]

    International Conference on Machine Learning , pages=

    3d infomax improves gnns for molecular property prediction , author=. International Conference on Machine Learning , pages=. 2022 , organization=

    work page 2022

  69. [69]

    2019 , journal =

    Strategies for Pre-training Graph Neural Networks , author =. 2019 , journal =

    work page 2019

  70. [70]

    2025 , journal =

    Platonic representation of foundation machine learning interatomic potentials , author =. 2025 , journal =

    work page 2025

  71. [71]

    International Conference on Learning Representations , year=

    Uni-mol: A universal 3d molecular representation learning framework , author=. International Conference on Learning Representations , year=

    work page

  72. [72]

    On representing chemical environments , author =. Phys. Rev. B , volume =

    work page

  73. [73]

    Randall Balestriero and Yann LeCun , year =

    work page

  74. [74]

    International Conference on Learning Representations , year =

    Pre-training via Denoising for Molecular Property Prediction , author =. International Conference on Learning Representations , year =. doi:10.48550/arXiv.2206.00133 , bibSource =

    work page doi:10.48550/arxiv.2206.00133

  75. [75]

    AI for Accelerated Materials Design - NeurIPS 2025 , year =

    LeMat-Traj: A Scalable and Unified Dataset of Materials Trajectories for Atomistic Modeling , author =. AI for Accelerated Materials Design - NeurIPS 2025 , year =

    work page 2025

  76. [76]

    International Conference on Machine Learning , year =

    Barlow Twins: Self-Supervised Learning via Redundancy Reduction , author =. International Conference on Machine Learning , year =

    work page

  77. [77]

    Scientific data , volume=

    Quantum chemistry structures and properties of 134 kilo molecules , author=. Scientific data , volume=. 2014 , publisher=

    work page 2014

  78. [78]

    Journal of chemical information and modeling , volume=

    Enumeration of 166 billion organic small molecules in the chemical universe database GDB-17 , author=. Journal of chemical information and modeling , volume=. 2012 , publisher=

    work page 2012

  79. [79]

    npj Computational Materials , volume=

    Benchmarking materials property prediction methods: the Matbench test set and Automatminer reference algorithm , author=. npj Computational Materials , volume=. 2020 , publisher=

    work page 2020

  80. [80]

    Physical review , volume=

    Self-consistent equations including exchange and correlation effects , author=. Physical review , volume=. 1965 , publisher=

    work page 1965

Showing first 80 references.