acceptodds
Under review as a conference paper at ICLR 2027

MOSAIC: Addressing Space for Local, Steerable, and Scalable Atomistic Generation

Abstract

Generative models are opening new possibilities in molecular discovery and materials science. Realizing this potential, however, requires both local control over atomic placement and the ability to assemble large, heterogeneous structures. We introduce MOSAIC, a hybrid mesh–particle representation that partitions space into anchored regions containing atom slots. Each slot encodes occupancy, atomic species, and a continuous displacement from its anchor, providing persistent spatial addresses for local conditioning, hierarchical generation, and scalable assembly. Because MOSAIC defines a representation rather than a model, it can be paired with different generative backbones. With the molecular backbone held fixed, its local support and coordinate prior improve molecular connectivity and validity; the resulting model samples efficiently across molecular sizes and enables fine-grained spatial control. At larger material scales, the same representation enables generation of atomistic polycrystals conditioned on coarse spatial layouts. Most notably, locality turns large-scale generation into a bounded assembly problem, enabling structure generation with more than one million atoms. Together, this establishes local spatial addressing as a practical foundation for steerable generation of matter across scales, from individual molecules to large-scale structures of broader engineering significance.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Generative Models, Molecular Generation, Multiscale Modeling, Physics Informed ML, Structure Prediction, Geometric Deep Learning

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_diPoOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents MOSAIC, a hierarchical generative modeling approach for molecule and materials design. By assigning atoms to local spatial slots while retaining continuous coordinates, MOSAIC aims to improve locality, controllability, and scalability. The approach is demonstrated on molecular generation and polycrystalline materials, including autoregressive assembly of systems containing more than one million atoms.

Critical strengths
  • The same representation is applied to rather different molecular and materials settings allowing some degree of generality.
  • Scalability to generation of polycrystalline structures with >1M atoms.
  • The molecular experiment shows that local support can substantially improve connectivity compared with a global-anchor baseline.
Critical weaknesses
  • The novelty relative to existing hierarchical spatial generators is insufficiently established. The paper itself identifies Uni-3DAR as the closest prior work, which already uses hierarchical spatial decomposition and fine-grained coordinates. MOSAIC appears to differ mainly in its mesh-anchored slot representation, continuous local displacements, and its use within diffusion/flow-based generation.
  • The manuscript could be simplified substantially. The relatively simple central idea is surrounded by considerable architectural and application-specific detail, confusing the main contributions.
  • Molecular generation performance is not competitive with the strongest presented baseline: MOSAIC reaches 77.2% connected-valid versus 84.7% for MegalodonFlow, with a particularly large gap in PoseBusters validity (54.1% vs. 89.8%). The latter points at fundamental issues with respect to downstream usability for molecular design.
  • The million-atom experiment demonstrates computational scalability, but provides less evidence for physical fidelity at that scale. More analysis of structural defects, relaxation stability, energies, or accumulation of autoregressive boundary errors would strengthen this result.
Questions for authors
  • What is the precise capability enabled by MOSAIC that cannot be achieved with Uni-3DAR or a similar hierarchical spatial representation? Given that Uni-3DAR is identified as the closest prior work, why is there no direct comparison?
  • Could the authors disentangle the contributions of hierarchical spatial support, continuous within-cell coordinates, multiple slots per cell, and the local coordinate prior?
  • Spatial conditioning seems to require substantial prior knowledge of the desired structure—for example, desired group centers for molecules or prescribed grain morphologies and orientations for polycrystals. How useful is this form of conditioning when the spatial arrangement itself is unknown and should be discovered?
  • How does error accumulate during autoregressive block assembly, and how physically stable are the resulting million-atom structures after relaxation?
4 comments
AuthorsOpenReview ↗

Author response to diPo

We would first like to thank you for taking the time to review our work. We appreciate your recognition of the improvements enabled by local support, the representation’s applicability across molecular and materials settings, and its computational scalability to systems containing over one million atoms. We address your concerns and questions point by point below.

W1. Novelty relative to Uni-3DAR and other hierarchical spatial generators

Thank you for highlighting the need to make this distinction clearer. The main methodology distinction is how MOSAIC and Uni-3dAR address the generation problem as the system size grows. Uni-3DAR compresses spatial structure into a token sequence and generates through an accumulating autoregressive history. MOSAIC instead assigns multiple continuous particles to coarse spatial addresses and jointly generates the current molecule or block.

The resulting advantages compared to Uni-3DAR are direct spatial control (key functionality) and bounded active context(memory efficient scaling). Spatial fields allow local editing and chemical-group placement. Such control is important for molecular inpainting in drug discovery (Schneuing et al., 2024) and positioning functional motifs in protein design (Watson et al., 2023). For largescale generation like the nickel assembly, each block uses bounded local context, keeping active neural memory independent of the assembled structure’s size. Uni-3dAR to our knowledge doesn’t show such largescale demonstrations. Our molecular implementation is also compact, 15.18M Mosaic versus 89.64M parameters for Uni-3dAR, with lower measured memory, latency and higher connected validity in the matched GEOM comparison (Fig F1). Also, we have now directly evaluated Uni-3DAR on 256 matched-count GEOM pairs using a common evaluator and the Figures are placed in Fig. F1 in the document supplied with this rebuttal. Our comparison scores every raw output from both methods, without the likelihood-based candidate ranking and selective retention used in Uni-3DAR’s reported protocol. MOSAIC achieves 89.1% connected-valid generation versus 81.3% for Uni-3DAR, a paired improvement of 7.81 percentage points. On the same MI300A GPU within each pair, the median paired Uni-3DAR/MOSAIC latency ratio is 1.37 for a fixed number of atoms, including both MOSAIC stages and the peak allocated memory is 225 versus 442 MiB. Uni-3DAR is faster for the smallest molecules. We will clarify these differences in the manuscript and add the direct comparisons shown in Fig F1 to the revised Supplementary Information.

  1. Schneuing, A., Harris, C., Du, Y., et al. (2024). Structure-based drug design with equivariant diffusion models. Nature Computational Science, 4(12), 899–909.
  2. Watson, J. L., Juergens, D., Bennett, N. R., et al. (2023). De novo design of protein structure and function with RFdiffusion. Nature, 620(7976), 1089–1100.

W2. Simplifying the presentation

Thank you for this helpful suggestion. We agree that the central idea should stand out more clearly. We will reorganize the method around this representation moving other domain-specific scientific details on architecture, nickel conditioning constructions and some other information to the appendix.

W3. Comparison and gap with MegalodonFlow

Thank you for highlighting this quality concern of the model. The purpose of the MegalodonFlow comparison is to contextualize the improvement achievable through spatial addressing and the purpose of the EDM derived backbone is to test how much an older, comparatively simple EDM-derived generator can improve through a simple change in spatial representation. MegalodonFlow provides a strong modern SOTA reference, explicitly co-generating coordinates, atom types, bonds and formal charges(unlike MOSAIC EDM). This comparison contextualizes the gains achievable through spatial addressing, while our matched-backbone ablation measures the contribution of local support and its coordinate prior directly. MegalodonFlow remains stronger on the reported molecular-quality endpoints, particularly PoseBusters, but MOSAIC nevertheless offers 3.07× connected-valid throughput under the tested schedules, together with spatial control. Its representation is independent of the particular molecular backbone used here.

AuthorsOpenReview ↗

Author response to diPo

W4. Physical fidelity of million-atom generation

Thank you for highlighting the need for stronger physical validation. We have now performed EAM potential energetics, all-mobile relaxation and boundary-resolved analysis on a 1,116,011-atom generated structure. After 3,000 relaxation (FIRE) iterations with the Ni99 potential, without deleting atoms, energy decreases from −1.978 to −4.290 eV/atom, FCC order increases from 26.32% to 46.05%, and the fraction of atoms with neighbours below 2 Å falls from 23.375% to 0.279%, approximately an 84-fold reduction. Minimization reached the iteration limit with residual contacts and forces. These results demonstrate substantial structural recovery; however we acknowledge that converged equilibrium and stability remain unestablished.

Q1. Novelty and capability relative to Uni-3DAR

Thank you for raising this question. We have addressed this in detail in W1.

Q2 Decomposing effects of support, continuous coordinates and prior.

Thank you for suggesting this clarification. The components serve distinct roles, which our experiments separate as follows: • Spatial support (existing ablation) Stage 1 allocates atom counts from coarse to fine while preserving the requested total. Our submitted matched EGNN comparison holds this hierarchy fixed: changing spatial support and its coordinate prior jointly raises connected validity from 24.3% to 65.9%. • Local prior (new): Using a shorter, matched training budget of 150,000 updates, we compare Gaussian and cell-uniform priors with identical predicted supports and matched coordinate variances. On 50 matched GEOM requests, they produce 27/50 and 25/50 connected-valid molecules, with nearly identical atom stability (81.20% and 81.25%). We therefore observe similar aggregate performance between the two prior families. • Continuous coordinates(new): On 1,000 generated molecules and 1,000 real GEOM conformers, rounding to 0.05 Å largely preserves validity, while 0.20 Å reduces it to 6.6% and 8.9%, respectively. This highlights the role of continuous offsets: the coarse mesh organizes space, while fine within-cell positions preserve bond geometry and molecular connectivity. • Multiple slots: Several atoms often occupy the same 3.2 Å cell. With only one slot per cell, only one could be represented and the remaining atoms would be missed. Multiple slots retain all these atoms. We will clarify these differences in the revised manuscript and add these additional ablations.

Q3 Importance of spatial conditioning and what happens without it?

Thank you for raising this critical fundamental question. In our molecular model, Stage 1 generates the spatial layout and local occupancies from atom count alone; therefore spatial prompts are optional for the molecular generation presented in the manuscript. However, our view is that fully unconditional sampling is not a sufficient objective for large, heterogeneous atomistic systems. Importantly, such information can be available before a complete atomic structure is known. Cryo-EM provides experimental three-dimensional electrostatic-potential maps of protein complexes, informing their shape and atomic organization (Pfab et al., 2021); X-ray diffraction-contrast tomography measures three-dimensional grain shapes and orientations (Ludwig et al., 2009); and phase-field simulations produce evolving morphology and orientation fields (Pusztai et al., 2005). We aim to complement progress in unconditional generation of small systems by addressing the spatial organization and controlled assembly needed for large, heterogeneous atomistic structures, something SOTA misses. We therefore regard generation conditioned on experimental or physically evolved spatial information as a central scientific objective, with the generator supplying just the unresolved atomic detail.

  1. Pfab, J., Phan, N. M., & Si, D. (2021). DeepTracer for fast de novo cryo-EM protein structure modeling and special studies on CoV-related complexes. Proceedings of the National Academy of Sciences, 118(2), e2017525118.
  2. Ludwig, W., Reischig, P., King, A., Herbig, M., Lauridsen, E. M., Johnson, G., Marrow, T. J., & Buffière, J.-Y. (2009). Three-dimensional grain mapping by X-ray diffraction contrast tomography and the use of Friedel pairs in diffraction data analysis. Review of Scientific Instruments, 80(3), 033905.
  3. Pusztai, T., Bortel, G., & Gránásy, L. (2005). Phase field theory of polycrystalline solidification in three dimensions. Europhysics Letters, 71(1), 131–137.
AuthorsOpenReview ↗

Author response to diPo

Q4 Error accumulation during block assembly.

Reciprocal descriptors repeatedly estimated from generated predecessors become unreliable. In our generated 1X8X8 block diagnostic, replacing them with reference-derived reciprocal fields raises FCC from 13.36% to 31.75% and reduces PTM ordered-fraction error from 0.474 to 0.278 (Fig. F 2a,b). The reported million-atom assemblies instead use persistent prescribed grain fields, with phases translated consistently between block origins. An audit of these fields gives 86.22° mean phase inconsistency without correction versus numerical precision with correction (Fig. F2c). This avoids repeated descriptor estimation; generated boundary atoms remain autoregressive inputs. Across three million-atom assemblies, each containing 3,375 blocks, we find no consistent monotonic FCC collapse with generation order. Local structural errors nevertheless remain concentrated near block interfaces. We will add these ablations in the revised version along with Figure F2 in the supplementary material.

Reviewer_5D4QOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The author/s propose/s MOSAIC, which is a representation for atomistic generation in which space is split into anchored cells, and each cell holds slots that store an atom type or an empty marker plus a continuous offset from the anchor. They use it for molecule generation on QM9 and GEOM-Drugs, and for building nickel polycrystals with over a million atoms block by block from coarse grain maps

Critical strengths
  • Very good core idea and good experimental design.
Critical weaknesses
  • million atom results are weak on quality. With causal descriptors the model reaches less than a quarter of the reference FCC fraction, a good share of the prescribed orientation states are missing in the large maps. There is no comparison with simply filling the map with oriented lattices.

  • it is unclear what MOSAIC adds for molecules. Uni-3DAR (Lu et al., Unified Cross-Scale 3D Generation and Understanding via Autoregressive Modeling), which the authors call closest to their hierarchy, is not compared, and the ablation’s global control is worse than released EDM on every shared metric and does not use equivariant OT coupling (Klein et al., Equivariant flow matching)

Questions for authors

-could you rerun the ablation with a stronger global control, using equivariant OT coupling and a retuned update bound and prior width, and also compare against Uni-3DAR with your evaluators?

-What FCC fraction and orientation-state recovery do you get for the source nickel itself under the same analysis? How do your structures compare with the same map filled with oriented lattices and relaxed?

4 comments
AuthorsOpenReview ↗

Authors response to 5D4Q

We sincerely thank you for the time and care devoted to reviewing our work. We appreciate your recognition of its strengths and your constructive comments, which have helped us clarify the contribution and strengthen the experimental evidence. We address your concerns point by point below.

W1. Million atom nickel results quality.

a. Causal-descriptor degradation. Thank you for highlighting this result provided in the supplementary. 13.36% FCC value concerns a 64-block continuation diagnostic in which the model estimates reciprocal descriptors from previously generated atoms. Those atoms need not contain the grain orientations required in the next region, making this a substantially different task from generation conditioned on a prescribed grain network. The reported million-atom assemblies receive persistent prescribed grain fields, with consistent phase translation between blocks. Their raw FCC fractions are 23.15%, 26.32% and 19.06%. For the primary 1,116,011-atom structure, subsequent EAM relaxation increases FCC from 26.32% to 46.05%. We will distinguish these protocols explicitly in the revision: the low causal-descriptor result diagnoses unreliable inference of the next region’s conditioning, whereas the main million-atom experiments evaluate generation under supplied spatial conditioning. Additional experimentation is provided in Figure F2 in the document provided with the rebuttal.

b. Missing orientation states. The three maps recover 26/41, 30/42 and 30/44 prescribed states. For the primary 42-state map, the 12 unrecovered states occupy 1.21% of the prescribed volume. Thus, nominal state counts disproportionately emphasize small regions. We will report this volume information alongside state recovery, while retaining fragmentation as a limitation. c. Comparison with oriented lattice. To address this reviewer concern we separately fine-tuned the Ni denoiser for masked reconstruction of smaller GB regions, adding a mask-conditioning adapter and occupancy objectives. The surrounding crystalline host is supplied and held fixed for a realistic comparison with similar oriented lattice. This however differs from the original million-atom assembly model and its all-mobile relaxation. For two completed validation assemblies, this model achieves mean relaxed, atom-count-adjusted energies 50.1 and 17.7 eV below the simple oriented-lattice filling, with atom-count deviations of +1.06% and −1.30% and better geometric coverage.

We will clarify the distinct conditioning protocols and include the lattice comparison, reporting raw and relaxed quality separately from the fine-tuned GB reconstruction results.

AuthorsOpenReview ↗

Authors response to 5D4Q

**W2. Comparison with Uni-3DAR, Global vs Local Ablation clarification and equivariant coupling. **

Comparison with Uni-3DAR - We have now directly evaluated Uni-3DAR on 256 matched-count GEOM pairs using a common evaluator and the Figures are placed in Fig. F1 in the document supplied with this rebuttal. Our comparison scores every raw output from both methods, without the likelihood-based candidate ranking and selective retention used in Uni-3DAR’s reported protocol. MOSAIC achieves 89.1% connected-valid generation versus 81.3% for Uni-3DAR, a paired improvement of 7.81 percentage points. On the same MI300A GPU within each pair, the median Uni-3DAR/MOSAIC latency ratio is 1.37 for a fixed number of atoms in FP32, including both MOSAIC stages and the peak allocated memory is 225 versus 442 MiB. Uni-3DAR is faster for the smallest molecules. The central methodological distinction is how generation scales with system size. Uni-3DAR encodes spatial structure as a token sequence and generates through an expanding autoregressive history. MOSAIC instead assigns multiple particles to coarse spatial addresses, retains their continuous coordinates, and jointly generates the current molecule or local block. This enables direct spatial control and memory-efficient assembly: at fixed local block settings, the structure can grow while the active neural context remains bounded.

Global versus local ablation. The global arm is our EGNN-based flow-matching implementation, trained with the same backbone, training budget and sampling protocol as the local arm. It serves as a controlled comparison of spatial support and coordinate priors, rather than a reproduction of released EDM. Introducing local anchors and their coordinate prior raises connected validity from 24.3% to 65.9%. Importantly, even this local ablation exceeds released EDM’s 34.9% under the shared GEOM evaluator. The molecular improvement therefore also holds against the released baseline itself.

Equivariant OT coupling. We conducted additional training experiments using rotation/permutation-aware global coupling based on the approximation in Klein et al. (2023), Eq. 16, alongside within-cell assignment for local support. These reduce squared coordinate matching costs by 75.3% globally and 33.3% locally.

To test molecular quality separately, we compared otherwise matched local Gaussian models trained with and without within-cell OT, keeping reference-derived coarse supports fixed. Both generated 4/16 connected-valid molecules. Thus, OT substantially improves geometric matching, but we observed no aggregate validity increase in this experiment. This is consistent with Klein et al. (2023), whose alanine-dipeptide experiments show that shorter transport paths need not improve sampling quality. OT optimizes noise–target pairing, while spatial support supplies locations and occupancies; the two mechanisms can be combined.

We will include the direct Uni-3DAR comparison and clarify the matched ablation, distinguishing spatial support from OT coupling.

Q1 Stronger global control with OT and Uni-3DAR comparison. As detailed in W2, we have added a direct Uni-3DAR comparison using our evaluator (Fig. F1) and new OT training experiments.

New experimentMetricResult
Global rotation/permutation-aware OTReduction in squared coordinate matching cost75.3%
Within-cell local OTReduction in squared coordinate matching cost33.3%
Matched local Gaussian models, OT off vs. onConnected-valid molecules4/16 vs. 4/16

We will report the coupling-cost diagnostics separately from molecular-quality measurements and state the scope of the completed comparisons explicitly.

AuthorsOpenReview ↗

Authors response to 5D4Q

Q2 State Recovery and comparison with oriented lattice.

Under the same PTM RMSD cutoff of 0.2, the full periodic nickel source has 81.78% FCC, and the matched open 240 Å source window has 79.16% FCC. We also re-estimated orientations from 125 retained source regions with atomic halos, recovering 41/41, 42/42 and 43/44 prescribed states for maps A–C. This measures detectability within selected source regions, rather than independent whole-volume grain-network recovery. We performed the requested same-map oriented-lattice comparison using a common 3,000-iteration Ni99 EAM relaxation budget, without atom deletion. MOSAIC’s FCC fraction increases from 26.32% to 46.05%, reaching −4.290 eV/atom; lattice filling reaches 84.76% FCC and −4.381 eV/atom. Both calculations reach the iteration limit. This demonstrates substantial structural recovery while identifying a remaining fidelity gap at million-atom scale. We also performed a separate fine-tuning study focused on reconstructing grain-boundary regions within a known crystalline host which is much closer to the simple oriented lattice structure. Starting from the existing nickel generator, we mask the GB region and train the model to generate its atom occupancies and continuous coordinates while preserving the surrounding host. Conditioning is computed from the visible host atoms. We compare these reconstructions with oriented-lattice filling under the same fixed-host EAM relaxation protocol, allowing only the reconstructed GB atoms to move. Compared with cleaned oriented-lattice filling, the fine-tuned model lowers the relaxed energy above the perfect-FCC nickel reference by 50.1 and 17.7 eV. It also reduces the fraction of GB grid points lying more than 2 Å from any atom from 0.169% to 0.064% and 0.164% to 0.097%, respectively.

Reviewer_jC3ZOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper offers hybrid mesh-particle based representation approach for crystal generation. The authors compare their model's performance on small molecules to different approaches.

Critical strengths

The paper is trying to tackle an important issue for molecular representation. Which is usually left out in the current SOTA push for frontier materials science research.

Critical weaknesses

The main weaknesses are:


  1. The experiment design is weak and doesn’t necessarily highlight the work’s novelty and contribution. For example, QM9 results does not support in the favor of work’s direction (Table 1) as well as the amorphous-silicon (Table 21) and solid-solid interface (Table 22). CDVAE and DiffCSP are significantly better than the proposed work (about 10 times). Author's say 'Because MOSAIC defines a representation rather than a model, it can be paired with different generative backbones'. However, there is no significant experiment and results highlighting it.

  2. The atom size is too small while calling their title ‘scalable’. Figure 2 shows the atom size of maximum of 120 and the molecular quality is 26% at that range. Figure 11 is only representation of generation without accuracy compared to other models but only their results hence does not show any improvement over other models.

Questions for authors
  1. QM9 is mainly a property prediction dataset rather than crystal generation. Why only used 10000 molecules from QM9 when it consists of 130K+ molecules? Could you justify this selection? Any why not use datasets like MP-20 [1] or Perov-5 [2] which are pretty much main benchmark dataset in this field?

  2. The experimental results does not really support the author's models benefit. Can author's cleafify why we should use their model when its way weaker than CDVAE and DiffCSP in multiple dimensions. These models are currently not even state of the art. There are better and newer models in this field and could authors compare to them as well [3, 4, 5]?

  3. The 'scalable' is in title and there is no significant results for scalability. Could author's provide details for their scalability? If possible show experimental results compared to state-of-the-art (MatterGen, CDVAE, DiffCSP, FlowMM [6]) in terms of memory, speed, and accuracy?

  4. How it is really connected to engineering and new technology? The results and the experimental design doesn't really highlight. For example, a hydrogen storage or catalyst special analyze could have been done.

  5. There is no discussion of scopes and the limitations of the approach. Could authors provide details about this?

  6. Could authors clearfiy how MOSAIC can be paired with different generative backbones and what are its benefits? The current experimental design does not support in the direction.

[1] Jain, A., Ong, S. P., Hautier, G., Chen, W., Richards, W. D., Dacek, S., ... & Persson, K. A. (2013). Commentary: The Materials Project: A materials genome approach to accelerating materials innovation. APL materials, 1(1). [2] Castelli, I. E., Landis, D. D., Thygesen, K. S., Dahl, S., Chorkendorff, I., Jaramillo, T. F., & Jacobsen, K. W. (2012). New cubic perovskites for one-and two-photon water splitting using the computational materials repository. Energy & Environmental Science, 5(10), 9034-9043. [3] Ruple, L., Torresi, L., Schopmans, H., & Friederich, P. (2026). Symmetry-aware Bayesian flow networks for crystal generation. npj Computational Materials, 12(1), 182. [4] Park, H., & Walsh, A. (2026). Guiding generative models to uncover diverse and novel crystals via reinforcement learning. Nature Machine Intelligence, 1-13. [5] Zeni, C., Pinsler, R., Zügner, D., Fowler, A., Horton, M., Fu, X., ... & Xie, T. (2025). A generative model for inorganic materials design. Nature, 639(8055), 624-632. [6] Miller, B. K., Chen, R. T., Sriram, A., & Wood, B. M. (2024). Flowmm: Generating materials with riemannian flow matching. arXiv preprint arXiv:2406.04713.

4 comments
AuthorsOpenReview ↗

Authors response to jC3Z

Thank you for taking the time to review our work and for raising these concerns. We first clarify the scientific problem and intended scope of the paper. The review’s framing of our work primarily as another crystal-structure-prediction model overlooks its broader scientific objective: controlled atomistic construction across scales. We are asking how generative modeling can construct heterogeneous systems whose spatial organization, interfaces and local structural requirements are part of the design specification. Our choice of demonstrations follows this objective. Molecular benchmarks assess precise spatial control at small scales, including chemical-group placement and local editing, with broader relevance to drug discovery and protein design [1,2]. The nickel demonstrations address large-scale atomistic assembly, guided by coarse spatial information about the desired structure. Such information can come from experimental characterization or larger-scale simulations, with the atomistic generator supplying the missing microscopic detail. Polycrystalline grain-boundary networks, thin films, nanocomposites, porous materials and biomolecular assemblies illustrate the need to connect atomic detail with larger-scale spatial organization [3,4]. These demonstrations therefore examine complementary aspects of the same problem: precise local control and organized construction at extended scales. MP20-style crystal prediction addresses a different objective; optimizing conventional CSP benchmark performance is not the aim of this paper.

  1. Schneuing, A., et al. (2024). Structure-based drug design with equivariant diffusion models. Nature Computational Science, 4, 899–909.
  2. Watson, J. L., et al. (2023). De novo design of protein structure and function with RFdiffusion. Nature, 620, 1089–1100.
  3. Chen, L.-Q. (2002). Phase-field models for microstructure evolution. Annual Review of Materials Research, 32, 113–140.
  4. Ludwig, W., et al. (2009). Three-dimensional grain mapping by X-ray diffraction contrast tomography and the use of Friedel pairs in diffraction data analysis. Review of Scientific Instruments, 80, 033905.

We now address the reviewer concerns point by point below :-

W1. Novelty of MOSAIC and compatibility with other architectures.

a. The novelty lies in the representation and the capabilities it enables. Uni-3DAR is the closest hierarchical comparison, which we now evaluate directly in Fig. F1(shared as zenodo link below). MOSAIC gives continuous particles persistent spatial addresses, enabling spatial organization during generation: chemical-group steering in molecules and field-conditioned assembly of extended atomistic systems. These capabilities are not adequately assessed by rankings on QM9 or small periodic crystals alone. We deliberately use an older, comparatively simple EDM-derived EGNN backbone to test how much local spatial addressing contributes. Holding the backbone and training protocol fixed, local support and its coordinate prior raise connected validity from 24.3% to 65.9%. The full implementation reaches 77.2% versus released EDM’s 34.9% under the shared GEOM evaluator. MegalodonFlow remains stronger on molecular-quality endpoints, but jointly predicts coordinates, atom types, bonds and formal charges. Our contribution is an additional representation-level capability that can be incorporated into stronger generators.

b. Compatibility with different architectures is already demonstrated, using an EDM-derived equivariant EGNN for molecules and a distinct GeoFiLM graph–Transformer architecture for nickel. We additionally implemented local spatial addressing in SEMLA(Irwin et al., 2025) , retaining its coordinate, atom, bond and charge predictions. On 10,000 validation requests, local SEMLA conditioned on reference-derived coarse occupancies produces connected-valid molecules with at least 80% occupancy overlap in 72.19% of cases, versus 8.72% for count-only global SEMLA after rigid alignment. This demonstrates spatial control on another backbone. We acknowledge the weaker silicon and interface benchmark results as limitations of those implementations. They do not negate the controlled molecular gains or the additional spatial-control and assembly capabilities. Moreover, the supplementary materials experiments test the breadth of the representation across qualitatively different physical regimes: molecular structures, chemical disorder in Fe–Ni–Cr alloys, amorphous silicon, multicomponent solid–solid interfaces and extended nickel polycrystals. These involve different bonding, spatial organization and boundary conditions, yet share the same local occupancy-and-continuous-coordinate representation, using separately trained models. Irwin, R., Tibo, A., Janet, J. P., & Olsson, S. (2025). SemlaFlow – Efficient 3D Molecular Generation with Latent Attention and Equivariant Flow Matching. AISTATS, Proceedings of Machine Learning Research, 258, 3772–3780.

AuthorsOpenReview ↗

Authors response to jC3Z

W2. Scalability with MOSAIC representation.

The scalability claim concerns extended atomistic generation, not a maximum molecule size of 120 atoms. MOSAIC generates nickel polycrystals containing a million atoms within a 240 Å × 240 Å × 240 Å domain, 24 nm on each side. At fixed local generation settings, active neural context remains bounded as the structure grows. We distinguish this computational scalability from physical fidelity, which we assess through structural metrics and EAM relaxation. Figure 2 evaluates molecular quality as size increases under a controlled comparison. At 120 atoms, MOSAIC produces 26% connected-valid molecules versus 0% for released EDM under the same evaluator. The relevant result is the improvement over the baseline at that size; 120 atoms is not the framework’s generation limit. More broadly, local support and its coordinate prior raise connected validity from 24.3% to 65.9% with the same EGNN backbone and training protocol. The full MOSAIC implementation reaches 77.2%, compared with 34.9% for released EDM and 84.7% for MegalodonFlow. Thus, spatial addressing substantially improves an older EDM-derived generator, bringing it closer to a stronger modern model on connected validity while enabling spatial control and extended assembly.

**Q1. Benchmarking on MP20 and CSP datasets **

We chose QM9 to test how well MOSAIC generates small molecules in 3D. This is an established use of QM9 and Geom-Drugs datasets alongside property prediction, as shown by most other molecular generation models. We trained both stages on the same 100,000 QM9 molecules, then drew 10,000 new samples from the trained model for evaluation(and not for training). That evaluation sample count follows earlier work. Hoogeboom et al., 2022, Xu et al., 2023, Joshi et al., 2025. For materials, we asked whether MOSAIC could take a map of grain shapes and orientations and place atoms across an entire microstructure. We chose million-atom nickel polycrystals because they let us measure how closely the generated structure follows the prescribed grain map and examine the atomic structure at grain boundaries. MP-20 and Perov-5 address a different task: generating small periodic unit cells. Xie et al., 2022.

Q2. Comparison with SOTA

We directly compared MOSAIC with released EDM and MegalodonFlow, a leading baseline for de novo 3D molecular generation, on 1,000 matched GEOM-Drugs requests using a shared evaluator. MOSAIC achieves 77.2% connected validity, compared with 34.9% for EDM and 84.7% for MegalodonFlow. MegalodonFlow has the higher PoseBusters pass rate among valid molecules (89.8% vs. 54.1%), while MOSAIC produces 3.07× as many connected-valid molecules per second under the models’ native sampling schedules. For materials, the reviewer is comparing MOSAIC against methods designed for a different objective. Periodic crystal design is not directly relevant to the objective of this paper. CDVAE, DiffCSP, MatterGen, SymmBFN, and related methods generate or discover periodic unit cells. MOSAIC instead addresses conditional atomistic construction of large, nonperiodic microstructures from prescribed grain shapes and orientations, including a nickel polycrystal spanning approximately 240 Angstrom per side and containing over one million atoms. A stronger periodic-crystal score would not answer whether a model can reconstruct atomic structure across a prescribed grain network or resolve its interfaces. We therefore evaluate MOSAIC using same-map oriented-lattice filling and physical relaxation controls, which directly test fidelity for our target task.

Q3. Scalability of MOSAIC representation.

Thank you for raising this. The scalability claim and the relevant memory, latency, and million-atom results are addressed in W2, including the direct Uni-3DAR comparison and the distinction between periodic crystal generation and MOSAIC’s large-scale conditional microstructure assembly.

AuthorsOpenReview ↗

Authors response to jC3Z

Q4. Connection to engineering and emerging technologies.

Thank you for raising this point. Hydrogen storage and catalyst-specific studies are separate application domains and are not the focus of this work. Our materials target is grain-boundary engineering, where atomic-scale boundary structure and chemistry directly affect corrosion, fracture, creep, diffusion, irradiation response, and electrical and thermal transport. Grain-boundary engineering is an established materials-engineering strategy for controlling these properties (Cantwell et al., 2020; Sutton and Balluffi, 1995).

A second application area is large amorphous and thin-film organic systems. The morphology, molecular orientation, and local packing of organic semiconductor films influence charge transport, stability, and OLED performance (Dalal et al., 2015; Yu et al., 2021). On the molecular side, spatially controlled generation and editing also aligns with structure-guided biomolecular and protein-design tasks, where geometry and functional-group placement are part of the design objective.

• Cantwell, P. R., Frolov, T., Rupert, T. J., Krause, A. R., Marvel, C. J., Rohrer, G. S., Rickman, J. M., & Harmer, M. P. (2020). Grain boundary complexion transitions. Annual Review of Materials Research, 50, 465–492. • Sutton, A. P., & Balluffi, R. W. (1995). Interfaces in crystalline materials. Clarendon Press, Oxford. • Dalal, S. S., Walters, D. M., Lyubimov, I., de Pablo, J. J., & Ediger, M. D. (2015). Tunable molecular orientation and elevated thermal stability of vapor-deposited organic semiconductors. Proceedings of the National Academy of Sciences, 112(14), 4227–4232. • Chen, X.-L., Tao, X.-D., Wei, Z., Meng, L., Lin, F.-L., Zhang, D.-H., Jing, Y.-Y., & Lu, C.-Z. (2021). Thermally activated delayed fluorescence amorphous molecular materials for high-performance organic light-emitting diodes. ACS Applied Materials & Interfaces, 13(39), 46909–46918.

Q5. Scope and limitations.

Thank you for asking us to clarify the scope. MOSAIC is designed for atomistic generation with spatial organization, local editing, and large-system assembly. Its current demonstrations cover molecular generation and nickel polycrystals conditioned on coarse grain morphology and orientation fields. The nickel results establish million-atom generation, while also exposing the main current limitation. Raw structures retain substantial interface errors and incomplete FCC order, and relaxation improves but does not fully close the gap to oriented-lattice filling. Improving physical fidelity at large scale remains an important direction.

The representation also raises broader questions about how atomistic detail can be generated from spatial information available before a complete structure exists. Relevant extensions include multicomponent alloys and complex grain-boundary networks conditioned on composition, stress, or orientation fields, amorphous and thin-film organic materials where morphology and molecular packing control device behavior, and biomolecular systems where spatial maps or functional-site constraints guide molecular editing. Future work should test whether MOSAIC can learn these conditioning fields directly, combine them with stronger molecular or materials backbones, and produce structures that remain physically stable after relaxation and dynamics.

Q6. Compatability with different backbones

As addressed in W1, MOSAIC is a representation layer that maps atoms to spatially anchored slots containing occupancy, species, and continuous coordinates. These states can be processed by different generative architectures. We demonstrate this with an EDM-derived equivariant network for molecules, a GeoFiLM graph–Transformer for materials, and a new SEMLA-based molecular experiment. Across these backbones, MOSAIC adds explicit spatial conditioning, local particle-number control, and bounded-context assembly.

Reviewer_Hi1gOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper introduces MOSAIC, a representation designed to improve atomistic generation. It partitions space into anchored cells, each containing slots that encode occupancy, atomic species, and continuous displacements from the cell’s anchor. For molecules, generation combines a count hierarchy with an equivariant flow over these slots. For nickel polycrystals, the method uses a local graph–Transformer flow incorporating morphology maps and reciprocal-space phase features, with autoregressive assembly enabling the generation of structures containing approximately 1.1 million atoms.

Critical strengths
  1. The representation claim is isolated. With CountDiT, architecture, and training held fixed, replacing a global center and Gaussian prior by local anchors and a cell-wise uniform prior lifts the GEOM-Drugs connected-valid rate from 24.3% to 65.9% at nearly the same seconds per molecule.

  2. The transport argument is simple and relevant: expected per-particle action along the anchor-relative path is at most h2, independent of atom count at fixed cell width, and permutation ambiguity drops from Θ(N log N) to at most N log K. The size sweep matches that story: at 120 atoms EDM’s connected-valid rate is 0% and MOSAIC’s is 26%, inside the training range.

  3. Spatial prompts are wired through both stages. Raising stage-1 guidance from 0 to 2 moves requested-group F1 from 0.491 to 0.667 while preserving the requested atom count.

  4. The materials half shows that the same slot state can carry a prescribed grain map. Reciprocal-phase features cut local structural errors by 5.2–35× in a controlled nickel setting, and the autoregressive procedure produces 1.09–1.12 million-atom polycrystals from coarse maps.

Critical weaknesses
  1. On the matched GEOM-Drugs evaluator, MOSAIC beats EDM (connected-valid 77.2% versus 34.9%, conditional PoseBusters 54.1% versus 19.0%) and loses to MegalodonFlow (84.7% and 89.8%). The 3.07× connected-valid throughput versus MegalodonFlow is a speed–quality trade, and MegalodonFlow also emits bonds and charges while MOSAIC reconstructs them for scoring. The abstract’s claim of improved molecular connectivity and validity is easy to read as a broader state-of-the-art claim.

  2. Main-text evidence for million-atom generation is mostly visual, plus grain-label recovery (2/2, 3/3, and 8/8 on small maps). Boundary drift, energy, and defect statistics after many autoregressive steps are deferred to the appendix. Grain-state recovery on a coarse map is a weak test of crystallinity.

Questions for authors
  1. After 24 nm of autoregressive assembly, what is the block-wise FCC fraction, RDF error, and energy relative to the seed block, as a function of raster distance from the seed?

  2. If stage 2 is replaced by a MegalodonFlow-class backbone on the same slots, does local support still help, or was the 24.3% to 65.9% gain specific to this EGNN?

  3. What happens to the connected-valid rate if the requested size exceeds the training maximum, where the h2 bound would be most interesting?

2 comments
AuthorsOpenReview ↗

Authors response to Hi1g

We thank the reviewer for the careful assessment and for recognizing the controlled representation ablation, transport argument, and spatial conditioning. Your questions motivated additional analyses of physical fidelity, backbone transfer, and molecular size extrapolation.

W1. Molecular quality and the abstract

Thank you for pointing this out. We will revise the abstract to state that the molecular result is an improvement over the EDM-derived baseline and a matched representation ablation, rather than a broad state-of-the-art validity claim. MegalodonFlow remains stronger on the reported molecular-quality endpoints.

W2 and Q1. Million-atom fidelity and accumulated error

We have analyzed FCC versus generation order across all three 24 nm assemblies. After excluding the supplied seed and external surface layer, we observe no common monotonic FCC collapse. For Map B, the first and last generation-order deciles have 26.9% and 26.7% FCC. The clearer deficit occurs at block interfaces. The assemblies use persistent grain fields with phases translated consistently between block origins. This avoids repeatedly estimating reciprocal conditioning from generated predecessors, a source of degradation identified by our diagnostic in Fig. F2. Generated boundary atoms still provide context and can propagate atomistic errors. We also evaluated whole-system nickel EAM relaxation of 1,116,011 atoms, allowing every atom to move without deleting atoms. After 3,000 iterations, energy decreases from −1.978 to −4.290 eV/atom, FCC increases from 26.3% to 46.0%, and the fraction of atoms with nearest neighbors below 2 Å decreases from 23.38% to 0.279%. We will bring the essential physical-quality results into the main text.

Q2. Transfer to a stronger molecular backbone

We performed a new SEMLA-based (Irwin et al., 2025) backbone experiment, retaining its native coordinate, atom-type, bond, and formal-charge predictions. On 10,000 requests, connected-valid generation with at least 80% overlap with the requested cell occupancies reaches 72.19% with local support versus 8.72% for the matched global model, after the same rigid-alignment procedure. Local generation receives reference-derived coarse occupancies, while global generation receives atom count. Connected validity alone is 82.19% versus 83.44% on this cohort. Thus, the experiment demonstrates transfer of spatial control to another architecture, while the large connectivity gain observed with EGNN remains backbone-dependent in our tested implementations.

Irwin, R., Tibo, A., Janet, J. P., & Olsson, S. (2025). SemlaFlow – Efficient 3D Molecular Generation with Latent Attention and Equivariant Flow Matching. AISTATS, Proceedings of Machine Learning Research, 258, 3772–3780.

Q3. Generation beyond the training maximum

We tested count-only generation near and beyond the 181-atom training maximum. Full MOSAIC produces 1/12 connected-valid molecules at 176 atoms and 0/12 at 192 atoms, preserving the requested count in every output. The h2 bound concerns expected local coordinate transport, while connected molecular generation additionally requires consistent chemistry across cells. Reliable extrapolation beyond the training range remains an empirical limitation that we will state explicitly.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.