pith. sign in

arxiv: 2604.23921 · v1 · submitted 2026-04-27 · 💻 cs.LG · cs.AI

Crystal structure prediction using graph neural combinatorial optimization

Stavros Gerolymatos , J. Kyle Brubaker , Martin J. A. Schuetz , Vladimir V. Gusev This is my paper

Pith reviewed 2026-05-08 04:46 UTC · model grok-4.3

classification 💻 cs.LG cs.AI
keywords crystal structure predictiongraph neural networkscombinatorial optimizationexpander graphsGumbel-Sinkhornatom allocationmaterials discovery
0
0 comments X

The pith

Graph neural networks allocate atoms on discrete positions to predict crystal structures, outperforming heuristics and competing with commercial solvers.

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

The paper introduces a graph neural network method for crystal structure prediction by solving the combinatorial problem of placing atoms on a fine grid to minimize energy. It builds expander graphs to connect positions and capture interactions, then applies a Gumbel-Sinkhorn layer to ensure the generated structures have the correct chemical makeup. This unsupervised sampling approach generates feasible low-energy configurations. Readers would care if it works because discovering new crystal structures drives advances in technology through better materials, and this method uses available GPU power to handle bigger problems than before. The results show it beats standard heuristic methods and keeps up with a commercial solver on various compositions.

Core claim

We demonstrate a neural combinatorial optimization method for crystal structure prediction using graph neural networks to sample feasible atomic structures unsupervised. Leveraging expander graphs over discrete positions to model short- and long-range interactions and the Gumbel-Sinkhorn approach to enforce stoichiometry, the technique produces structures that outperform classical heuristic approaches and compete with a commercial optimization solver across different chemical compositions.

What carries the argument

Graph neural network operating on expander graphs of discrete atomic positions, with a Gumbel-Sinkhorn layer to enforce stoichiometry.

If this is right

  • Scaling crystal structure prediction to larger and more complex systems becomes feasible by utilizing expanding GPU resources.
  • Computational costs for CSP decrease compared to exact optimization methods for large instances.
  • Exploration of atomic configuration space is possible without symmetry constraints or labeled data.
  • Integration into broader materials discovery pipelines accelerates the identification of new functional materials.

Where Pith is reading between the lines

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

  • This method might extend to optimizing other discrete placement problems in chemistry, such as molecular packing.
  • Combining the GNN sampler with more accurate energy models could improve the reliability of predicted structures.
  • Testing on periodic systems with defects or interfaces could reveal limitations in the current graph construction.
  • The competitive performance suggests potential for hybrid approaches with traditional solvers for even better results.

Load-bearing premise

The expander graphs over discrete positions capture the essential short- and long-range atomic interactions needed for the neural network to distinguish low-energy structures.

What would settle it

If the method is applied to a set of well-known crystal compositions and consistently fails to find structures with energies matching or better than known ground states within the solver's time frame, the performance claim would be falsified.

Figures

Figures reproduced from arXiv: 2604.23921 by J. Kyle Brubaker, Martin J. A. Schuetz, Stavros Gerolymatos, Vladimir V. Gusev.

Figure 1
Figure 1. Figure 1: GNT-CSP: neural combinatorial approach to CSP. A. Following Gusev et al. [2023], a dense grid of points is defined within a unit cell as potential atomic positions. We seek to identify the optimal, in terms of energy, allocation of the required number of atoms corresponding to a chosen stoichiometry (in this figure, SrTiO3 ), to this set of positions. B. A computational graph based on a classic expander gr… view at source ↗
Figure 2
Figure 2. Figure 2: ROG curves for the configurations predicted by each method as a function of the combinatorial search space (in logarithmic scale) of the investigated compositions. The orange curve corresponding to SA is always below the blue curve of greedy but above the green curve of the traditional solver used in IPCSP. Therefore, IPCSP performs consistently better than both classical methods SA, while SA outperforms g… view at source ↗
Figure 3
Figure 3. Figure 3: Number of epochs our method needed to compute the lowest energy structure of each composition as a function of their combinatorial search space (in logarithmic scale). Our model tends to need more training epochs to find low-energy feasible structures as the the combinatorial search space of a composition gets larger, while inherent symmetries found in SrTiO3 with Z = {8, 27} require a comparable number of… view at source ↗
Figure 4
Figure 4. Figure 4: Barbell graph. One typical example of a bottleneck that can lead to oversquashing can be found in a barbell graph, where the green edge is responsible to propagate information between two large groups of nodes. 18 view at source ↗
Figure 5
Figure 5. Figure 5: The developed model architecture for our GNN. An embedding block is used to create a d0-dimensional node embedding for each of the n positions. We create b0-dimensional edge embeddings by taking the respective columns of the interactions matrix Q and passing them from a linear layer. Node and edge embeddings are fed to GINE layers which generate node embeddings according to message-passing as shown in the … view at source ↗
Figure 6
Figure 6. Figure 6: Instances of one-hop node connectivities for radius cutoff, Margulis 3D and Gabber-Galil 3D graphs. Each row of subfigures illustrates the connections (blue edges) of the same central node (in yellow color) with its one-hop neighbors (in green color) for all three graphs. Radius cutoff graphs capture mainly connections between neighboring nodes, while periodicity introduces few connections between more dis… view at source ↗
Figure 7
Figure 7. Figure 7: ROG curves for the structures GNT-CSP computed with respect to type of graph we performed the message-passing on and as a function of the combinatorial search space, in logarithmic scale, of the compositions we experimented with. The blue curve that corresponds to the case of the Gabber-Galil 3D graph is below both the orange one (3D margulis) and the green one (radius cutoff), indicating the consistently … view at source ↗
read the original abstract

Crystalline materials are widely used in technological applications, yet their discovery remains a significant challenge. As their properties are driven by structure, crystal structure prediction (CSP) methods play a central role in computational approaches aiming to accelerate this process. Previously, CSP has been approached from a combinatorial optimization perspective, with the core challenge of allocating atoms on a fine grid of predefined discrete positions within a unit cell while minimizing their interaction energy. Exact mathematical optimization methods provide guaranteed solutions, but they become computationally expensive for large-scale instances, where the atomic configuration space grows rapidly, particularly in the absence of additional symmetry constraints. In this work, we introduce a neural combinatorial optimization approach to the atom allocation challenge and, subsequently, CSP, based on graph neural networks (GNNs), which can effectively sample from the distribution of feasible structures in an unsupervised manner. We leverage expander graphs to construct computational graphs over discrete positions that capture both short- and long-range interactions between atoms, and employ the Gumbel-Sinkhorn approach to enforce the desired stoichiometry of the generated structures. We demonstrate that our method outperforms classical heuristic approaches and is competitive with a commercial optimization solver across a range of chemical compositions. This enables the use of ever-expanding GPU infrastructure to tackle the inherent combinatorial challenges of CSP, paving the way for scaling beyond current capabilities.

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 / 3 minor

Summary. The paper introduces a neural combinatorial optimization method for crystal structure prediction (CSP) by framing atom allocation on a discrete grid as a combinatorial problem. It constructs expander graphs over positions to model short- and long-range interactions, employs Gumbel-Sinkhorn layers to enforce stoichiometry, and uses a GNN to sample feasible low-energy structures in an unsupervised manner. The central claim is that this approach outperforms classical heuristics and is competitive with a commercial solver across chemical compositions, enabling GPU scaling for larger CSP instances.

Significance. If the benchmarks hold, the work offers a meaningful advance by integrating GNN message passing with combinatorial techniques (expander graphs and Gumbel-Sinkhorn) to address the exponential growth of configuration space in CSP. The unsupervised sampling and explicit motivation for long-range encoding via expanders are strengths that could complement existing ML-CSP pipelines and leverage expanding GPU resources.

major comments (2)
  1. [§3.1] §3.1 (expander graph construction): The claim that expander graphs capture both short- and long-range interactions sufficiently for feasible low-energy sampling is load-bearing for the outperformance result; the manuscript should include a controlled comparison (e.g., against k-nearest-neighbor graphs or full-interaction baselines on small cells) to quantify the contribution of the expansion property.
  2. [Results section] Results section, benchmark tables: The competitiveness with the commercial solver is central to the main claim, yet the manuscript must report exact solver time limits, hardware, and whether the GNN sampling budget was matched to the solver's; without these controls the comparison risks being non-equivalent.
minor comments (3)
  1. [Abstract] Abstract: Add at least one quantitative metric (e.g., success rate or energy difference) and mention of error bars or number of compositions to make the outperformance claim verifiable from the summary.
  2. [Notation] Notation: Define the discrete position set and stoichiometry vector consistently across equations; the current presentation occasionally overloads symbols for grid points and atom types.
  3. [Figure 2] Figure 2 (or equivalent): The diagram of the GNN + Gumbel-Sinkhorn pipeline would benefit from an explicit legend distinguishing message-passing edges from the Sinkhorn projection step.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive feedback and for recognizing the potential of integrating GNNs with combinatorial optimization techniques for crystal structure prediction. We address each major comment below and have prepared revisions to enhance the manuscript's rigor and transparency.

read point-by-point responses

  1. Referee: [§3.1] §3.1 (expander graph construction): The claim that expander graphs capture both short- and long-range interactions sufficiently for feasible low-energy sampling is load-bearing for the outperformance result; the manuscript should include a controlled comparison (e.g., against k-nearest-neighbor graphs or full-interaction baselines on small cells) to quantify the contribution of the expansion property.

    Authors: We agree that an explicit ablation would better substantiate the role of the expansion property. In the revised manuscript we will add a controlled comparison on small cells, evaluating expander graphs against k-nearest-neighbor graphs and dense (full-interaction) baselines. The new experiments will report sampling success rates and achieved energies, thereby quantifying the contribution of long-range connectivity provided by the expander construction. revision: yes

  2. Referee: Results section, benchmark tables: The competitiveness with the commercial solver is central to the main claim, yet the manuscript must report exact solver time limits, hardware, and whether the GNN sampling budget was matched to the solver's; without these controls the comparison risks being non-equivalent.

    Authors: We acknowledge that precise experimental controls are necessary for a fair comparison. We will revise the results section and benchmark tables to state the exact wall-clock time limits and hardware platform used for the commercial solver. We will also document the GNN sampling budget (number of forward passes and temperature schedule) and confirm that it was calibrated to be comparable in total compute to the solver's allotted time, thereby ensuring equivalence of the reported performance. revision: yes

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The paper introduces a new unsupervised neural combinatorial optimization procedure for CSP: GNN message passing on expander graphs over discrete positions, combined with Gumbel-Sinkhorn for stoichiometry enforcement. Performance is evaluated empirically against classical heuristics and a commercial solver on multiple compositions. No load-bearing step reduces by construction to a fitted parameter, self-definition, or self-citation chain; the expander-graph construction and training objective are explicitly specified as independent design choices motivated by interaction range, and the reported outperformance is not claimed as a first-principles derivation but as algorithmic benchmarking. The method is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach rests on standard assumptions from graph neural networks and combinatorial optimization; no new physical axioms or invented particles are introduced.

free parameters (1)
  • GNN architecture hyperparameters
    Number of layers, hidden dimensions, learning rate, and temperature parameters for Gumbel-Sinkhorn chosen or tuned during training.
axioms (1)
  • domain assumption Expander graphs adequately encode short- and long-range atomic interactions for the purpose of structure sampling
    Invoked when constructing the computational graph over discrete positions.

pith-pipeline@v0.9.0 · 5545 in / 1129 out tokens · 49026 ms · 2026-05-08T04:46:13.929772+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

4 extracted references · 4 canonical work pages

  1. [1]

    Stefano Curtarolo, Gus L

    doi:10.1021/cr2003272. Stefano Curtarolo, Gus L. W. Hart, Marco Buongiorno Nardelli, Natalio Mingo, Stefano Sanvito, and Ohad Levy. The high-throughput highway to computational materials design.Nature Materials, 12(3):191–201, 2013. ISSN 1476-4660. doi:10.1038/nmat3568. URLhttps://doi.org/10.1038/nmat3568. Anubhav Jain, Shyue Ping Ong, Geoffroy Hautier, W...

    work page doi:10.1021/cr2003272 2013

  2. [2]

    Vladimir V Gusev, Duncan Adamson, Argyrios Deligkas, Dmytro Antypov, Christopher M Collins, Piotr Krysta, Igor Potapov, George R Darling, Matthew S Dyer, Paul Spirakis, et al

    doi:10.4230/LIPIcs.MFCS.2022.8. Vladimir V Gusev, Duncan Adamson, Argyrios Deligkas, Dmytro Antypov, Christopher M Collins, Piotr Krysta, Igor Potapov, George R Darling, Matthew S Dyer, Paul Spirakis, et al. Optimality guarantees for crystal structure prediction.Nature, 619(7968):68–72, 2023. OI Pulkkinen, P Gautam, V Mustonen, and T Aittokallio. Multiobj...

    work page doi:10.4230/lipics.mfcs.2022.8 2022

  3. [3]

    Naval Re- search Logistics Quarterly2(1-2), 83–97 (1955).https://doi.org/https://doi

    doi:https://doi.org/10.1002/nav.3800020109. URL https://onlinelibrary.wiley.com/doi/abs/10. 1002/nav.3800020109. Gurobi Optimization, LLC. Gurobi Optimizer Reference Manual, 2024. URLhttps://www.gurobi.com. Vijay V . Vazirani.Approximation Algorithms. Springer Publishing Company, Incorporated, 2010. ISBN 3642084699. Ilhem Quenel Boussaid, Julien Lepagnot,...

    work page doi:10.1002/nav.3800020109 2024

  4. [4]

    PoseNet: A convolutional network for real-time 6-dof camera relocalization,

    Curran Associates Inc. Diederik P. Kingma and Max Welling. Auto-Encoding Variational Bayes. In2nd International Conference on Learning Representations, 2014. George Papandreou and Alan L. Yuille. Perturb-and-map random fields: Using discrete optimization to learn and sample from energy models. InProceedings of the 2011 International Conference on Computer...

    work page doi:10.1109/iccv 2014