Pith. sign in

REVIEW 4 major objections 5 minor 16 references

Efficient and Scalable Density Functional Theory Hamiltonian Prediction through Adaptive Sparsity

T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read SPHNet learns to prune most tensor products in Hamiltonian prediction and still outperforms QHNet.

desk verdict Real speedups and honest ablations, but the accuracy advantage over QHNet is likely inflated by a residual-target change the baselines don't use. read the letter →

arxiv 2502.01171 v2 pith:HIV5JXCF submitted 2025-02-03 cs.LG physics.comp-ph

Erpai Luo , Xinran Wei , Lin Huang , Yunyang Li , Han Yang , Zaishuo Xia , Zun Wang , Chang Liu
show 2 more authors
Bin Shao Jia Zhang
This is my paper · ORCID
classification cs.LGphysics.comp-ph
keywords HamiltonianmatrixpredictionSE(3)equivariantneuralnetworkadaptivesparsitytensorproductpruningClebsch-Gordancoefficientsdensityfunctionaltheorymolecularpropertyefficientdeeplearning
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

SPHNet is an attempt to remove the computational bottleneck in machine-learned prediction of the Kohn-Sham Hamiltonian, the matrix from which molecular energies, orbital energies, and wavefunctions can be read off. The bottleneck is the Clebsch-Gordan tensor product: an SE(3) equivariant network must combine atomic features of many angular-momentum orders, and the cost grows quadratically with atom count and steeply ($O(L^6)$) with the maximum order $L$. The paper's claim is that most of those interactions are redundant and can be pruned adaptively. SPHNet learns two gates—one that keeps only important atom pairs, one that keeps only important cross-order tensor-product channels—and freezes the selection after a short warm-up, reporting Hamiltonian mean absolute errors lower than QHNet on QH9 and PubChemQH while training 3.3x to 7.1x faster and using 25% to 30% of the GPU memory. If the sparsity assumption transfers, larger molecules and larger basis sets become tractable for equivariant Hamiltonian models.

What carries the argument

The load-bearing objects are two learned gates plus a scheduler. The Sparse Pair Gate computes a per-pair weight $W^p_{ij}$ from zero-order features and inner products of a pair, then keeps only the top fraction of pairs under the scheduler. The Sparse Tensor Product Gate assigns a weight $W^{\ell_1,\ell_2,\ell_3}_c$ to each allowed angular-momentum combination $(\ell_1,\ell_2,\ell_3)$ in the Clebsch-Gordan tensor product and multiplies the path weights by those retained values. The Three-phase Sparsity Scheduler first selects uniformly at random so all combinations receive gradients, then for exactly one epoch selects by top weight, then freezes the selected set and the weight vector so the tensor products can be executed as static, optimized kernels. The machinery therefore turns a dense $O(N^2)$ pair set and an $O(L^6)$ per-tensor-product cost into a sparse, static computation graph whose size is set by the sparsity rate $k$.

What would settle it

Run SPHNet at 70% sparsity on PubChemQH with the scheduler's freeze epoch $t$ varied (for example 1, 3, 6, 10, and 20) while keeping the total training budget fixed; if Hamiltonian MAE improves substantially as $t$ increases, the fixed early selection is cutting out combinations that later training would rely on. A second check is to train with the gate weights frozen but the selected set re-randomized at the same sparsity rate; similar accuracy would mean the specific selected pairs and paths are not what carries the performance.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that a static, learned sparsity pattern can replace the fully connected set of tensor products in an equivariant Hamiltonian predictor without sacrificing accuracy. SPHNet's Sparse Pair Gate scores every atom pair with a sigmoid of a learned linear function of scalar features, and its Sparse Tensor Product Gate scores every angular-momentum triple $(\ell_1,\ell_2,\ell_3)$ with a learnable weight; a Three-phase Sparsity Scheduler selects by random draw during early training, switches to top-weight selection at epoch $t=3$, then freezes both the selection and the gate weights. The reported results are concrete: on QH9-stable iid the Hamiltonian MAE is $45.48\times10^{-6}\,\mathrm{E_h}$ versus $76.31$ for QHNet, and on PubChemQH it is $97.31$ versus $123.74\times10^{-6}\,\mathrm{E_h}$, with training speedups of $4.0\times$ and $7.1\times$ and memory at 30% and 25% of baseline. Ablations show accuracy stays nearly flat up to a size-dependent sparsity threshold—30% on MD17, 40% on QH9, 70% on PubChemQH—and that the retained atom pairs are increasingly long-range, suggesting the gate is selecting physically relevant interactions rather than simply short-distance neighbors.

Load-bearing premise

The load-bearing premise is that the pair and tensor-product weights learned by epoch 3, once frozen, pick out a static subset of atom pairs and angular-momentum combinations that remains near-optimal for the rest of training and at test time; if the important interactions shift as features develop, the early freeze would discard paths the network later needs, and the high-sparsity accuracies would not transfer.

Editorial extensions

If this is right

  • If the reported numbers hold, SPHNet makes Hamiltonian prediction practical for molecules around 3000 atomic orbitals on a single A6000 GPU, where the compared baseline is limited to roughly 1800.
  • The larger the basis set and molecule, the more redundancy there is to exploit: the observed speedup grows from about 3.3x to 4x on QH9 ($L_{\max}=4$) to 7.1x on PubChemQH ($L_{\max}=6$), so sparsity is most valuable exactly where tensor products are most expensive.
  • Setting the sparsity rate by system size—30% on small trajectories, 40% on QH9, 70% on PubChemQH—keeps Hamiltonian MAE near the dense model while buying most of the speedup, so practitioners can treat molecule size as a rough guide without per-model sweeps.
  • The same sparse gates improve the existing QHNet architecture by 3.3x when both gates are added, suggesting the sparsification is a transferable module rather than a property of SPHNet's own design.
  • Because retained pairs skew long-range, the model's learned graph is not a distance cutoff; long-range interactions are treated as important even though they are costly, which matters for charged or polar systems.

Reading between the lines

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

  • A direct stress test the authors did not run is to move the freeze epoch $t$ from 3 to later values; if MAE at 70% sparsity drops significantly, the reported accuracy depends on early freezing being benign rather than on sparsity itself being free.
  • The pair gate's preference for long-range pairs, if it reflects electrostatics, predicts that the gate's rank ordering should transfer across molecules with similar charge environments; that could be tested by training the gate on one chemical class and applying it to another.
  • The same Clebsch-Gordan path pruning should apply to any equivariant architecture that builds messages from tensor products, such as many-body message-passing models, but the speedup would be smaller if the architecture already factors the tensor product into cheaper operations.
  • Because the sparse tensor-product gate multiplies path weights before the tensor product, it changes the function realized by the network; an implicit assumption is that reweighting discarded paths to zero does not need a compensating adjustment elsewhere, and a test would be measuring how much the norm of retained weights drifts after freezing.
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 / 5 minor

Summary. The paper introduces SPHNet, an SE(3)-equivariant graph neural network for predicting DFT Hamiltonian matrices. The main novelty is a pair of learned sparse gates: a Sparse Pair Gate that selects a subset of atom pairs and a Sparse Tensor Product Gate that prunes cross-order Clebsch-Gordan combinations, together with a Three-phase Sparsity Scheduler that first trains with random selection, then adaptively selects and finally freezes the sparse structure. The authors report on QH9, PubChemQH, and MD17 datasets that SPHNet achieves state-of-the-art Hamiltonian MAE while being up to 7.1x faster in training than QHNet and reducing GPU memory by up to 75%, with ablations over sparsity rates, scheduler phases, and gate configurations.

Significance. If the accuracy claims survive controlled comparison, this is a practically valuable contribution: it targets a real computational bottleneck in equivariant Hamiltonian prediction and demonstrates a plausible mechanism to prune tensor-product computations, with measured near-linear speedups on a single GPU. The paper is unusually thorough in its component ablations, reports speed and memory on consistent hardware (single A6000), and releases code. The main caveats are that the headline accuracy comparison is confounded by a residual-target formulation and that the headline numbers lack error bars; the sparsity-rate selection also uses the same curves that define the reported accuracy. These issues are fixable and do not invalidate the efficiency contribution, but they do need to be addressed before the 'state-of-the-art accuracy' claim can be accepted.

major comments (4)
  1. [Appendix A.3 (Eq. 16), Appendix A.1] SPHNet predicts the residual target ΔH = H_ref − H_init, where H_init is a cheap minao initial guess, and the appendix states that the scale and variance of ΔH are roughly an order of magnitude smaller than those of H_ref on large datasets. However, Appendix A.1 says the QHNet baseline was run with its default setting, which is not described as using the same residual target or loss. Because Tables 1 and 2 report Hamiltonian MAE comparisons (e.g., 97.31 vs 123.74 on PubChemQH), the accuracy advantage attributed to adaptive sparsity may instead reflect the easier regression target. Please re-run the baselines with the identical ΔH target and loss, or report the baseline performance on the same residual target, before claiming state-of-the-art accuracy.
  2. [Section 5.4 (Fig. 3, Table 7)] The per-dataset sparsity rates used in the main experiments (0.1/0.3/0.4/0.7 for MD17, QH9, PubChemQH) are chosen as the 'critical turning points' read from the very same MAE-versus-sparsity curves (Fig. 3) that constitute the reported accuracy results. This means the reported accuracies at those rates are the result of selecting the hyperparameter on the evaluation curve, which can bias the headline numbers optimistically. The paper should state explicitly whether Fig. 3 shows validation or test MAE, and should report the full trade-off (as partially given in Table 9) alongside the main comparison tables so readers can see the accuracy cost at the chosen operating point.
  3. [Table 3 (Water row), Section 5.3] The MD17 water result is not 'comparable' as claimed: SPHNet achieves H MAE 23.18e-6 Eh and ϵ MAE 182.29e-6 Eh versus QHNet's 10.79e-6 Eh and 33.76e-6 Eh, which is approximately 2x and 5x worse, respectively. While the paper correctly frames MD17 as a secondary small-molecule setting, the text overstates the accuracy match. Please either temper the claim or provide an explanation for why the water molecule is an outlier in the otherwise comparable MD17 results.
  4. [Tables 1, 2, 3; Table 4] The main comparison tables report single-run results without error bars, while Table 4 shows substantial run-to-run variation (e.g., H MAE 97.31 ± 0.52 for the full scheduler but 122.79 ± 19.02 without the adaptive stage). Since some of the headline margins are modest (e.g., 97.31 vs 99.98 for WANet on PubChemQH), the 'state-of-the-art accuracy' claim is not statistically supported. Please report means and standard deviations over multiple seeds for at least the main table entries, and specify the number of seeds used.
minor comments (5)
  1. [Appendix B.4] The sentence 'the Sparse Pair Gate selects 30% of these combinations' appears to refer to the Sparse Tensor Product Gate, since the surrounding text is about tensor product combinations; please correct the terminology.
  2. [Section 4.2 (heading)] The heading 'Sparse Tenor Product Gate' contains a typo ('Tenor' should be 'Tensor').
  3. [Table 9] The reported memory usage is non-monotonic in sparsity: at 60% sparsity it is 5.12 GB/sample, while at 70% it rises to 5.62 GB/sample. If this reflects implementation details or variable batch shapes, it would be helpful to note that memory does not strictly decrease with sparsity.
  4. [Section 4.1, Table 7] The freeze epoch t=3 is fixed for all datasets and is not ablated. A sensitivity analysis over t, even on one dataset, would strengthen the claim that the scheduler's three phases are robust to the phase-switch timing.
  5. [Equation (12), Appendix D.4] The notation '⊗_{ℓ1,ℓ2}^{ℓ3}' is used in several equations but is not formally defined; please define it once (e.g., as the Clebsch-Gordan coupled tensor product) to avoid ambiguity with the symbol '⊗' used for the tensor expansion in Appendix D.6.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the sparse-gating accuracy and speed claims are evaluated against external DFT labels and the external QHNet baseline.

full rationale

The paper's central derivation is self-contained in the relevant sense: the sparse gate masks are learned from training data via Equations 3-8 and then frozen, and the reported Hamiltonian MAE, orbital-energy MAE, memory, and speed are measured against external DFT Hamiltonians (QH9, PubChemQH, MD17) and the external QHNet baseline. This is ordinary pruning evaluation, not a prediction of a fitted quantity. The sparsity rate is chosen from the ablation curve in Section 5.4 and the full rate-vs-MAE table is disclosed (Table 9), so reporting the selected operating point is transparent hyperparameter selection rather than a constructed result. The residual target ΔH = H_ref - H_init introduced in Appendix A.3 changes the regression difficulty and may confound the accuracy comparison with baselines run in their default setting, but that is a comparison-fairness or correctness concern, not circularity: the target transformation is not defined in terms of SPHNet's output. The comparison to WANet is limited by the paper's own statement that WANet has not been open-sourced, so those rows cannot be independently reproduced; that is a reproducibility limitation, not a circular derivation. The only self-referential elements are the adoption of the authors' prior long-short-range message-passing block (Li et al., 2024) and the WANet baseline from the same group (Li et al., 2025b); neither is used to justify the sparse-gating claim, no uniqueness theorem is imported, and no prior result is renamed, so there is no load-bearing self-citation chain.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The method is empirical, so the ledger is mostly about hyperparameters and assumptions used in evaluation. The per-dataset sparsity rates are the main free parameters because they determine both the reported speedup and the reported accuracy. The core modeling assumption is that top-k scalar weights identify interaction combinations that can be permanently discarded. No new physical entities are introduced.

free parameters (2)
  • per-dataset sparsity rate k = MD17 water 0.1, other MD17 0.3; QH9 0.4; PubChemQH 0.7
    Chosen in Section 5.4 as the largest sparsity before MAE degrades on each dataset. This directly controls both the reported speedup and the reported accuracy, so it is a free parameter in the headline claim.
  • scheduler phase switch epoch t = 3 for all datasets
    Set in Table 7 without a reported sensitivity study. It determines when top-k selection and freezing begin, and the frozen graph is a major source of the speedup.
assumptions (4)
  • domain assumption DFT ground-truth Hamiltonians in QH9, PubChemQH, and MD17 are reliable labels.
    All evaluation and the sparsity-rate selection in Section 5.4 assume the reference calculations (B3LYP and PBE with def2-SVP and def2-TZVP) are correct; no label error analysis is provided.
  • domain assumption The learned gate weights W_c and W_p reflect the importance of tensor-product combinations and pairs.
    Equations 4, 7, 8 and the analysis in Section 5.5 treat top-k by these scalar weights as selecting the most important interactions. The paper does not prove that this optimization criterion is the correct one.
  • ad hoc to paper The three-phase random, adaptive, fixed scheduler converges to a stable near-optimal sparse set.
    Section 4.1 and the ablation in Table 4 justify this empirically, but convergence is assumed from the design rather than derived or formally analyzed.
  • domain assumption A static, frozen computation graph is accuracy-equivalent to the adaptive sparse graph.
    Table 4 shows that removing the fixed stage keeps accuracy but reduces speedup from 7.09x to 5.45x, so the reported speedup depends on this implementation-level assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Efficient and Scalable Density Functional Theory Hamiltonian Prediction through Adaptive Sparsity." pith.science (2026). https://pith.science/paper/HIV5JXCF

@misc{pith2026250201171,
  author       = {Pith},
  title        = {Pith review of: Efficient and Scalable Density Functional Theory Hamiltonian Prediction through Adaptive Sparsity},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HIV5JXCF}},
  note         = {Machine review of arXiv:2502.01171}
}
read the original abstract

Hamiltonian matrix prediction is pivotal in computational chemistry, serving as the foundation for determining a wide range of molecular properties. While SE(3) equivariant graph neural networks have achieved remarkable success in this domain, their substantial computational cost--driven by high-order tensor product (TP) operations--restricts their scalability to large molecular systems with extensive basis sets. To address this challenge, we introduce SPHNet, an efficient and scalable equivariant network, that incorporates adaptive SParsity into Hamiltonian prediction. SPHNet employs two innovative sparse gates to selectively constrain non-critical interaction combinations, significantly reducing tensor product computations while maintaining accuracy. To optimize the sparse representation, we develop a Three-phase Sparsity Scheduler, ensuring stable convergence and achieving high performance at sparsity rates of up to 70%. Extensive evaluations on QH9 and PubchemQH datasets demonstrate that SPHNet achieves state-of-the-art accuracy while providing up to a 7x speedup over existing models. Beyond Hamiltonian prediction, the proposed sparsification techniques also hold significant potential for improving the efficiency and scalability of other SE(3) equivariant networks, further broadening their applicability and impact. Our code can be found at https://github.com/microsoft/SPHNet.

Figures

Figures reproduced from arXiv: 2502.01171 by the authors.

Figure 1
Figure 1. (A) The number of tensor products grows quadratically with the number of atoms N, as the Hamiltonian includes features for all possible atomic pair combinations. (B) The time cost of tensor products grows with the sixth power of their order L, where the increase in order corresponds to the expansion in the number of orbital types in the DFT basis set. For example, the def2-SVP basis set requires a maximum order of 4… view at source ↗
Figure 2
Figure 2. (A) The overall architecture of SPHNet. Atomic numbers and atomic coordinates are first passed through the Vectorial Node Interaction Blocks to obtain atomic features x ℓ i . Subsequently, the Sparse Pair Gate selects the key pair set (i, j) for the Spherical Node Interaction Blocks, where the irreps x ℓ i are elevated from the ℓ = 1 to Lmax during the interaction process. Next, the Sparse Tensor Product Gate in the… view at source ↗
Figure 3
Figure 3. The effect of different sparsity rates on the model per￾formance on the datasets with different scale of molecules. See the detailed computational cost scaling in Appendix B.2 of Pub￾ChemQH Dataset. Additionally, we observed that the threshold at which the MAE begins to rise significantly becomes larger as the com￾plexity of the molecular system increases. For the smallest MD17 system, the critical turning point is … view at source ↗
Figures from the paper (5 more)
Figure 5
Figure 5. Figure 5: Comparison of the training speed with the increasing size of Hamiltonian. 5.7. Additional Ablation study on SPHNet To evaluate the effectiveness of the Three-phase Sparsity scheduler and the Sparse Gate on the model performance, we carried out a series of ablation stud…
Figure 6
Figure 6. Figure 6: The time consumption for a single tensor product with scaling sparsity rates. B.4. Selected Combination Set Details Here we presented the details of the selected combination sets in all Sparse TP Gate in the SPHNet. We first analysis the output order of the selected co…
Figure 7
Figure 7. Figure 7: The proportions of each output order being selected in different tensor product gates. We then check the combination weights for each of the selected combinations. The details are shown in [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: The GPU memory consumption with scaling molecule size. Y l m =    √ i 2 Y l −|m| − (−1)mY l |m| , if m < 0. Y l 0 , if m = 0. √ 1 2  Y l |m| + (−1)mY l −|m|  , if m > 0. (20) To ensure both consistency and standardized treatment throughout the analysis, the Co…
Figure 9
Figure 9. Figure 9: The components of SPHNet. (A) The Vectorial Node Interaction Block, which uses a long-short range message-passing mechanism. (B) The Expansion block. (C) The Spherical Node Interaction Block. (D) The Diagonal block in the Pair Construction block. (E) The Non-Diagonal b…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

16 extracted references · 9 canonical work pages

  1. [6]

    and the weight calculation (Equation 9), are necessary operations in our framework, even without the sparse pairwise gate. E. Additional Related works E.1. SE(3) Equivariant Neural Network The SE(3) equivariant neural network is one of the most used models in the field of AI for chemistry (Fuchs et al., 2020; Du et al., 2022; Musaelian et al., 2023; Liao ...

    work page 2020

  2. [7]

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

    Liao, Y .-L., Wood, B., Das, A., and Smidt, T. Equiformerv2: Improved equivariant transformer for scaling to higher- degree representations. arXiv preprint arXiv:2306.12059,

    arXiv

  3. [8]

    and De Fabritiis, G

    Th¨olke, P. and De Fabritiis, G. Torchmd-net: Equivariant transformers for neural network based molecular poten- tials. arXiv preprint arXiv:2202.02541,

    arXiv

  4. [9]

    Towards flexible, efficient, and effective tensor product networks

    Wang, N., Lin, C., Bronstein, M., and Torr, P. Towards flexible, efficient, and effective tensor product networks. In NeurIPS 2023 Workshop: New Frontiers in Graph Learning,

    work page 2023

  5. [11]

    Infusing Self-Consistency into Density Functional Theory Hamiltonian Prediction via Deep Equilibrium Models

    Wang, Z., Liu, C., Zou, N., Zhang, H., Wei, X., Huang, L., Wu, L., and Shao, B. Infusing self-consistency into density functional theory hamiltonian prediction via deep equilibrium models. arXiv preprint arXiv:2406.03794 ,

    work page Pith review arXiv

  6. [12]

    (A) The Vectorial Node Interaction Block, which uses a long-short range message-passing mechanism

    The components of SPHNet. (A) The Vectorial Node Interaction Block, which uses a long-short range message-passing mechanism. (B) The Expansion block. (C) The Spherical Node Interaction Block. (D) The Diagonal block in the Pair Construction block. (E) The Non-Diagonal block in the Pair Construction block. D.1. RBF The RBF here refers to the Radial Basis Fu...

    work page 2022

  7. [14]

    Last, the non-diagonal feature fij is outputted after normalization and sph linear the same as in the Spherical Node Interaction block. There are two separate Pair Construction Blocks that receive atom representations from the two SO(3) Convolution Blocks respectively, and the final node pair feature is the addition of these two Pair Construction Blocks’ ...

    work page 2023

  8. [15]

    is an improved version of Equiformer, which scales effectively to higher-order representations by replacing SO(3) convolutions with efficient eSCN convolutions (Passaro & Zitnick, 2023), and outperforming the traditional network such as GemNet (Gasteiger et al.,

    work page 2023

Show all 16 references
  1. [16]

    presents an efficient method to perform SO(3) equivariant 22 Efficient and Scalable Density Functional Theory Hamiltonian Prediction through Adaptive Sparsity convolutions. It reduces the computational complexity by aligning node embeddings’ primary axis with edge vectors, tra...

    2021

  2. [2000]

    Geomformer: A general architecture for geomet- ric molecular representation learning

    Chen, T., Luo, S., He, D., Zheng, S., Liu, T.-Y ., and Wang, L. Geomformer: A general architecture for geomet- ric molecular representation learning. arXiv preprint arXiv:2406.16853,

    arXiv

  3. [2011]

    URL https://doi.org/ 10.1021/ct200412r

    doi: 10.1021/ct200412r. URL https://doi.org/ 10.1021/ct200412r. PMID: 26598153. Cances, E. and Le Bris, C. On the convergence of scf algo- rithms for the hartree-fock equations. ESAIM: Mathemat- ical Modelling and Numerical Analysis , 34(4):749–774,

    doi

  4. [2019]

    Se (3)-transformers: 3d roto-translation equivariant attention networks

    Fuchs, F., Worrall, D., Fischer, V ., and Welling, M. Se (3)-transformers: 3d roto-translation equivariant attention networks. Advances in neural information processing systems, 33:1970–1981,

    1970

  5. [2020]

    L., and Das, A

    Gasteiger, J., Shuaibi, M., Sriram, A., G ¨unnemann, S., Ulissi, Z., Zitnick, C. L., and Das, A. Gemnet-oc: devel- oping graph neural networks for large and diverse molecu- lar simulation datasets. arXiv preprint arXiv:2204.02782,

    arXiv

  6. [2022]

    URL https://doi.org/ 10.1021/acs.jctc.2c00509

    doi: 10.1021/acs.jctc.2c00509. URL https://doi.org/ 10.1021/acs.jctc.2c00509. PMID: 36136665. LeCun, Y ., Denker, J. S., and Solla, S. A. Optimal brain damage. In Advances in Neural Information Processing Systems (NeurIPS), pp. 598–605,

    doi

  7. [2023]

    Visnet: an equivariant geometry-enhanced graph neural network with vector- scalar interactive message passing for molecules

    11 Efficient and Scalable Density Functional Theory Hamiltonian Prediction through Adaptive Sparsity Wang, Y ., Li, S., He, X., Li, M., Wang, Z., Zheng, N., Shao, B., Liu, T.-Y ., and Wang, T. Visnet: an equivariant geometry-enhanced graph neural network with vector- scalar in...

    arXiv

  8. [2024]

    E2former: A linear-time efficient and equivariant transformer for scal- able molecular modeling

    Li, Y ., Huang, L., Ding, Z., Wang, C., Wei, X., Yang, H., Wang, Z., Liu, C., Shi, Y ., Jin, P., et al. E2former: A linear-time efficient and equivariant transformer for scal- able molecular modeling. arXiv e-prints, pp. arXiv–2501, 2025a. Li, Y ., Xia, Z., Huang, L., Wei, X.,...

    arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.