REVIEW 4 major objections 7 minor 4 cited by
Data-Driven Approach to Encoding and Decoding 3-D Crystal Structures
T0 review · 4 major / 7 minor · reviewed 2026-08-14 · deepseek-v4-flash
Pith's one-line read The paper claims that 3-D crystal structures can be compressed into a continuous latent space and decoded back into atomic positions accurate to within 0.5 Å for nearly 99% of atoms in single unit cells, with species identified correctly…
desk verdict A promising representation for crystals (density VAE + segmentation U-Net) with some real results, but the evaluation has enough unstated details that the headline accuracy number needs a careful look before you rely on it. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The central object is the density field $M_{i,j,k} = \frac{1}{\sigma^3(2\pi)^{3/2}} \sum_m Z_m \exp\left(-d(\vec{Z}_m, (i,j,k))^2 / 2\sigma^2\right)$ with $\sigma = 1.0$ Å, evaluated on a 30×30×30 grid over a 10 Å cube. This continuous field is the only representation the encoder and decoder see; discrete atom identities appear only in the species matrix used to supervise the segmentation network. The mechanism is joint training: the VAE loss $L_{\text{VAE}} = L_{\text{RE}} + \beta D_{\text{KL}} + \gamma L_{\text{BCE}}$ combines density reconstruction, a Kullback-Leibler term, and a weighted binary cross-entropy term from the U-Net segmentation, so the decoder is pushed to produce fields that segment cleanly into atoms. This coupling is what allows atom-level recovery from a representation that never explicitly stores atom identities.
What would settle it
Encode a crystal whose two adjacent atoms have atomic numbers Z and Z+1, then encode the same geometry with the two species swapped. The density fields from the Gaussian representation are nearly identical, so if species predictions on such mirrored inputs track the swap, the network is using information beyond the density field; if they do not, the two-thirds species ceiling is attributable to the representation itself.
Extended reading notes
Core claim
The central claim is that a voxelized density field, where each atom contributes an isotropic Gaussian of width $\sigma = 1.0$ Å scaled by its atomic number, is a workable learned representation for crystal structures. A variational autoencoder compresses this field into a 300-dimensional latent vector and reconstructs it, while a 3-D U-Net segments the reconstruction into atoms and assigns atomic numbers. The paper reports that for single unit cells nearly 99% of atoms are placed within 0.5 Å of their true location, nearly 90% of unit cells are reconstructed with the correct number of atoms, and 66% of species are correctly classified; when species are wrong, they are usually off by one or two atomic numbers. For repeating lattices, 50% of reconstructed atoms are within 0.2 Å of a true atom and 65.4% of predicted atoms within 0.5 Å receive the exact correct atomic number. The paper also demonstrates latent-space interpolation between two crystals and random draws from $z \sim \mathcal{N}(0,1)$ that produce plausible inter-atomic distances, although the decoded molecules are not yet accurate enough to relax in a density functional theory calculation.
Load-bearing premise
The load-bearing premise is that blurring every atom into a fixed-size Gaussian cloud whose height is its atomic number loses no information needed to recover both where the atom is and what element it is; atoms with nearly equal atomic numbers blur into almost identical clouds.
Editorial extensions
If this is right
- If crystals can be encoded and decoded as density fields, new material candidates can be proposed by sampling or optimizing in a continuous 300-dimensional latent space instead of enumerating explicit structures.
- Interpolating between two encoded crystals yields intermediate density fields that segment into atoms, suggesting that linear paths in latent space correspond to smooth structural transitions between compounds.
- Conditioning on the maximum density allows control over the heaviest atom present without moving the geometry, a step toward generating crystals with targeted elemental composition.
- The same encoder-decoder pipeline could apply to any 3-D density map, not only crystals, because it effectively learns a distance transform from an object.
- Random draws from the latent space produce inter-atomic distance distributions similar to real crystals, indicating the prior captures some physical constraints even though decoded molecules are not yet stable enough for density functional theory relaxation.
Reading between the lines
- The reported species accuracy of about two-thirds may be close to the information ceiling of the Z-scaled Gaussian representation, because atoms with adjacent atomic numbers create nearly identical density patterns; a model that outputs explicit atom identities or uses element-specific embeddings would likely recover species better.
- The 0.5 Å position accuracy is strong for geometry, but DFT relaxation requires both correct species and positions, so the method's present value is more as a search prior than as a final structure generator.
- Because the input cubes are randomly rotated but the architecture is not rotation-equivariant, the latent space may conflate orientation with geometry; an equivariant architecture could separate the two and make interpolation between crystals more physically meaningful.
- A testable extension suggested by the paper's setup is to add an auxiliary property-prediction head on the latent vector, conditioning generation on computed quantities such as band gap or formation energy.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes a data-driven approach to encoding and decoding 3-D crystal structures. It represents each crystal as a 30x30x30 voxelized density field built from isotropic Gaussians centered at atomic positions, trains a variational autoencoder to compress and reconstruct these density fields, and jointly trains a 3-D U-Net to segment the decoded density into atoms with predicted atomic numbers. Experiments are reported on single unit cells and on repeating lattices from the Materials Project, with claims of nearly 99% of atoms placed within 0.5 angstrom of their true positions in single cells, 65-66% exact species identification, and additional demonstrations of latent-space interpolation, random draws, and conditional generation.
Significance. If the reported accuracies hold under a correctly designed evaluation protocol, the paper would demonstrate a useful voxel-density VAE baseline for crystal structure encoding, with a clean formulation of the density representation in Eq. (1) and a sensible joint segmentation loss in Eq. (2). The central reconstruction task is well-defined and not circular, and the authors are explicit about the species-identification limitation. However, the quantitative claims are currently not established because of the ambiguous train/test split and the absence of a validation protocol, and the lack of any comparison baseline makes the contribution hard to position. As a proof-of-concept the paper is interesting; as a rigorous methods paper it needs substantial additional work.
major comments (4)
- [Section 3.1] The manuscript states 'We use 80% of the data for testing and the other 20% for training' and then describes sampling three random rotations for each single unit cell. It is not specified whether the 80/20 split is applied to the 46,744 unique crystal identifiers before rotation augmentation or to the augmented set of over 120,000 samples. If the split is applied after augmentation, rotated copies of the same unit cell appear in both training and test sets, so the reported 'nearly 99% of atoms placed within 0.5 Å' (Section 4.1) and the interpolation results would reflect memorization rather than generalization. Please clarify the protocol; if the split was after augmentation, the experiments must be rerun with a split on unique crystals performed before augmentation.
- [Section 3.2] No validation set is described, and the authors state that they 'experimented with γ = 0 and γ = 0.33 and found that 0.1 proved an acceptable intermediate.' Combined with the unusual 80% test / 20% training split, this raises the concern that the test set was used for hyperparameter selection, which would make all reported test accuracies optimistic. Please define a validation split, report validation-based selection of γ, β, latent dimension, and other hyperparameters, and confirm that the test set was used only once.
- [Sections 4.1 and 4.2.2] The central claim that 'nearly 99% of atoms are placed within 0.5 Å of their true location' is reported without error bars, without the number of random seeds, and without a precise definition of the metric (e.g., whether the threshold applies to both directions in Eq. (6) or only one). For repeating lattices, the text reports the 50th percentile at 0.2 Å, the 75th at under 1 Å, and the 90th at under 2 Å, which is a much weaker statement than 99% within 0.5 Å. Please provide the full cumulative distributions, error bars over seeds, and an exact definition of the single-cell metric.
- [Section 9] The discriminator validation described in Section 9 and Eq. (7) trains a network on linear interpolations between random latent draws and real encodings, both decoded by the same decoder, and then uses this network to score random draws from that same decoder. This is a circular test: it can only measure closeness to the decoder's own output manifold, not physical realism or agreement with the true crystal distribution. The claim that random samples 'obey a similar intra-molecule distance distribution' should be supported by direct comparison with the true distribution, with statistical tests, rather than by this discriminator score.
minor comments (7)
- [Title] The title contains a formatting artifact, 'DATA-D RIVEN'; please correct it.
- [Section 1] 'as aSMILES string' should read 'as a SMILES string'.
- [Section 2] There are several typos: 'start-of-the-art' should be 'state-of-the-art', 'hard constrains' should be 'hard constraints', and 'most be obeyed' should be 'must be obeyed'.
- [Section 3.1] The sentence 'We use 80% of the data for testing and the other 20% for training' is likely intended to be 80% training and 20% testing; please confirm and correct.
- [Section 4.2.2 and Fig. 4E] The text says the species comparison is 'independent of the distance to the nearest atom,' but the Fig. 4E caption restricts the comparison to atoms within 0.5 Å; please reconcile this inconsistency.
- [Section 4.2.2] The phrase '50% of all reconstructed atoms are in 0.2 Å' should read 'within 0.2 Å', and the later sentence switches to a 0.33 Å threshold for the 70% species-accuracy claim; please use consistent thresholds and define them clearly.
- [References] One reference is incomplete: 'Chen et al.).' appears in the paragraph on disentanglement; please provide the full citation.
Circularity Check
Core VAE reconstruction is self-contained; only the random-draw plausibility check is mildly circular because its discriminator is trained on the same decoder it validates.
-
other
[Section 9 (Accuracy of random draws in a repeating lattice), Eq. (7), Figs. 12-13]
"To test whether random samples from the latent space, z∼N (0, 1), decode to physically realistic molecules, we trained a discriminator. ... We then randomly draw z∼N (0, 1) from our trained network and pass this reconstruction D˜ into the discriminator network, which outputs a prediction of the distance from a true crystal reconstruction versus a random draw from a latent space of a previously trained network."
The discriminator used as evidence of physical plausibility is trained on outputs of the same decoder D and the same latent prior N(0,1) that generate the samples. Its training labels are constructed by interpolating between random normal latents and real encoded latents and decoding through D (Eq. 7). Therefore a high score for random draws only means those draws are close to real encoded latents under a metric learned from D's own output manifold; it is not an independent physical-validity test. This is a circular validation of the generative claim, though the paper also provides an external inter-atomic spacing comparison and hedges the claim, so it does not undermine the central reconstruction results.
full rationale
The central derivation is self-contained: the paper trains a VAE to reconstruct voxelized density fields and a U-Net to segment the decoder output into atoms, then evaluates reconstruction on a separate test set against ground-truth crystal structures. No load-bearing claim reduces by construction to its inputs; the density representation (Eq. 1) is a fixed preprocessing choice, not fitted to the evaluation metric, and the species matrix S is built from true atom positions and used as supervised labels. The reported position accuracy and species accuracy are measured against held-out data, and the paper explicitly acknowledges that species identification is imperfect and that the generated structures are not yet stable enough for DFT relaxation. There is no uniqueness theorem imported from the authors' prior work and no ansatz smuggled in via self-citation. The only mild circularity is the Section 9 discriminator-based validation of random draws, where the validator is trained on the same decoder it is used to assess; this is auxiliary rather than central and is supplemented by a direct comparison of inter-atomic spacing distributions. The possible train/test leakage from rotation augmentation is not established by the text and is a data-split concern rather than a circularity.
Assumptions & free parameters
free parameters (8)
- sigma (Gaussian width in density representation) =
1.0 Å
- voxel grid resolution =
30x30x30 over 10 Å (0.33 Å/voxel)
- latent dimension =
300
- beta (KL weight) =
not stated explicitly (beta/10 used in ablation)
- gamma (segmentation loss weight) =
0.1
- max unit-cell side length cutoff =
10 Å
- learning rate / batch size =
1e-5 / 24
- train/test split =
20% train / 80% test
assumptions (5)
- domain assumption DFT-relaxed crystal structures in the Materials Project are accurate ground-truth representations of stable crystals.
- ad hoc to paper Atomic electron density can be represented by a sum of isotropic Gaussian functions with a single width sigma = 1 Å, without element-specific widths or bonding effects.
- ad hoc to paper A 30x30x30 voxel grid at 0.33 Å resolution is fine enough to resolve atom positions within the reported tolerances.
- domain assumption A Gaussian prior p(z)=N(0,1) combined with a beta-weighted KL term produces a latent space worth sampling.
- domain assumption Connected components of the argmax segmentation output correspond one-to-one to atoms.
Cite this review
Pith. "Pith review of Data-Driven Approach to Encoding and Decoding 3-D Crystal Structures." pith.science (2026). https://pith.science/paper/63LIFUON
@misc{pith2026190900949,
author = {Pith},
title = {Pith review of: Data-Driven Approach to Encoding and Decoding 3-D Crystal Structures},
year = {2026},
howpublished = {\url{https://pith.science/paper/63LIFUON}},
note = {Machine review of arXiv:1909.00949}
}
read the original abstract
Generative models have achieved impressive results in many domains including image and text generation. In the natural sciences, generative models have led to rapid progress in automated drug discovery. Many of the current methods focus on either 1-D or 2-D representations of typically small, drug-like molecules. However, many molecules require 3-D descriptors and exceed the chemical complexity of commonly used dataset. We present a method to encode and decode the position of atoms in 3-D molecules from a dataset of nearly 50,000 stable crystal unit cells that vary from containing 1 to over 100 atoms. We construct a smooth and continuous 3-D density representation of each crystal based on the positions of different atoms. Two different neural networks were trained on a dataset of over 120,000 three-dimensional samples of single and repeating crystal structures, made by rotating the single unit cells. The first, an Encoder-Decoder pair, constructs a compressed latent space representation of each molecule and then decodes this description into an accurate reconstruction of the input. The second network segments the resulting output into atoms and assigns each atom an atomic number. By generating compressed, continuous latent spaces representations of molecules we are able to decode random samples, interpolate between two molecules, and alter known molecules.
Figures
Figures from the paper (12 more)
Forward citations
Cited by 4 Pith papers
-
A Periodic Bayesian Flow for Material Generation
CrysBFN adapts Bayesian Flow Networks to periodic crystal coordinates via von Mises distributions and entropy conditioning, achieving SOTA generation and 100x faster sampling.
-
Inverse Design of Amorphous Materials with Targeted Properties
A diffusion-based generative model (AMDEN) with energy-based Hamiltonian Monte Carlo refinement generates amorphous glass structures with targeted properties and low-energy relaxed states that standard denoising cannot reach.
-
Kinetic Langevin Diffusion for Crystalline Materials Generation
KLDM runs the diffusion process for crystal coordinates in Euclidean velocity space via left-trivialized kinetic Langevin dynamics on a torus, and reports competitive or state-of-the-art performance on CSP and DNG benchmarks.
-
Transformer-Enhanced Variational Autoencoder for Crystal Structure Prediction
TransVAE-CSP replaces the encoder in the CDVAE crystal generator with an equivariant transformer and per-dataset radial basis functions, improving reconstruction and generation on three benchmark datasets.
Reference graph
Works this paper leans on
-
[1]
Learning representa- tions and generative models for 3d point clouds
Panos Achlioptas, Olga Diamanti, Ioannis Mitliagkas, and Leonidas Guibas. Learning representa- tions and generative models for 3d point clouds. arXiv preprint arXiv:1707.02392,
-
[6]
Batch normalization: Accelerating deep network training by reducing internal covariate shift
Sergey Ioffe and Christian Szegedy. Batch normalization: Accelerating deep network training by reducing internal covariate shift. arXiv preprint arXiv:1502.03167,
-
[8]
Hyunjik Kim and Andriy Mnih. Disentangling by factorising. arXiv preprint arXiv:1802.05983,
-
[9]
Adam: A method for stochastic optimization
Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980,
-
[11]
Grammar variational autoen- coder
Matt J Kusner, Brooks Paige, and Jos ´e Miguel Hern ´andez-Lobato. Grammar variational autoen- coder. In Proceedings of the 34th International Conference on Machine Learning-Volume 70 , pages 1945–1954. JMLR. org,
work page 1945
-
[13]
Molecular geometry pre- diction using a deep generative graph neural network
Elman Mansimov, Omar Mahmood, Seokho Kang, and Kyunghyun Cho. Molecular geometry pre- diction using a deep generative graph neural network. arXiv preprint arXiv:1904.00314,
arXiv 1904
-
[14]
HoloGAN: Unsupervised learning of 3D representations from natural images
Thu Nguyen-Phuoc, Chuan Li, Lucas Theis, Christian Richardt, and Yong-Liang Yang. Holo- gan: Unsupervised learning of 3d representations from natural images. arXiv preprint arXiv:1904.01326,
work page Pith review arXiv 1904
-
[16]
URL http://distill.pub/2016/ deconv-checkerboard
doi: 10.23915/distill.00003. URL http://distill.pub/2016/ deconv-checkerboard. Ozan Oktay, Jo Schlemper, Loic Le Folgoc, Matthew Lee, Mattias Heinrich, Kazunari Misawa, Kensaku Mori, Steven McDonagh, Nils Y Hammerla, Bernhard Kainz, et al. Attention u-net: Learning where to look for the pancreas. arXiv preprint arXiv:1804.03999,
arXiv 2016
Show all 21 references
-
[17]
Generating diverse high-fidelity images with vq-vae-2
Ali Razavi, Aaron van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. arXiv preprint arXiv:1906.00446,
1906 arXiv
-
[18]
Tackling climate change with machine learning
David Rolnick, Priya L Donti, Lynn H Kaack, Kelly Kochanski, Alexandre Lacoste, Kris Sankaran, Andrew Slavin Ross, Nikola Milojevic-Dupont, Natasha Jaques, Anna Waldman-Brown, et al. Tackling climate change with machine learning. arXiv preprint arXiv:1906.05433,
1906 arXiv
-
[19]
uni-freiburg.de/Publications/2015/RFB15a
URL http://lmb.informatik. uni-freiburg.de/Publications/2015/RFB15a. (available on arXiv:1505.04597 [cs.CV]). Kristof T Sch ¨utt, Huziel E Sauceda, P-J Kindermans, Alexandre Tkatchenko, and K-R M ¨uller. Schnet–a deep learning architecture for molecules and materials. The Jour...
2015 arXiv
-
[20]
Generative models for automatic chemical design
Daniel Schwalbe-Koda and Rafael G ´omez-Bombarelli. Generative models for automatic chemical design. arXiv preprint arXiv:1907.01632,
1907 arXiv
-
[21]
Pointflow: 3d point cloud generation with continuous normalizing flows
Guandao Yang, Xun Huang, Zekun Hao, Ming-Yu Liu, Serge Belongie, and Bharath Hariha- ran. Pointflow: 3d point cloud generation with continuous normalizing flows. arXiv preprint arXiv:1906.12320,
1906 arXiv
-
[2008]
Graphnvp: An invert- ible flow model for generating molecular graphs
Kaushalya Madhawa, Katushiko Ishiguro, Kosuke Nakago, and Motoki Abe. Graphnvp: An invert- ible flow model for generating molecular graphs. arXiv preprint arXiv:1905.11600,
1905 arXiv
-
[2013]
Junction tree variational autoencoder for molecular graph generation
Wengong Jin, Regina Barzilay, and Tommi Jaakkola. Junction tree variational autoencoder for molecular graph generation. arXiv preprint arXiv:1802.04364,
-
[2014]
Auto-encoding variational bayes
Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114,
-
[2015]
Crystalgan: Learning to discover crystallographic structures with generative adversarial networks
Asma Nouira, Jean-Claude Crivello, and Nataliya Sokolovska. Crystalgan: Learning to discover crystallographic structures with generative adversarial networks. arXiv preprint arXiv:1810.11203,
-
[2016]
Large scale gan training for high fidelity natural image synthesis
Andrew Brock, Jeff Donahue, and Karen Simonyan. Large scale gan training for high fidelity natural image synthesis. arXiv preprint arXiv:1809.11096,
-
[2017]
De- factor: Differentiable edge factorization-based probabilistic graph generation
Rim Assouel, Mohamed Ahmed, Marwin H Segler, Amir Saffari, and Yoshua Bengio. De- factor: Differentiable edge factorization-based probabilistic graph generation. arXiv preprint arXiv:1811.09766,
-
[2018]
Generative and discriminative voxel modeling with convolutional neural networks
Andrew Brock, Theodore Lim, James M Ritchie, and Nick Weston. Generative and discriminative voxel modeling with convolutional neural networks. arXiv preprint arXiv:1608.04236,
-
[2019]
Work in progress
Preprint. Work in progress. Niklas W A Gebauer, Michael Gastegger, and Kristof T Sch¨utt. Symmetry-adapted generation of 3d point sets for the targeted discovery of molecules. arXiv preprint arXiv:1906.00957,
1906 arXiv
Reviewed August 14, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.