pith. sign in

arxiv: 2606.28578 · v1 · pith:WXEFPOJQnew · submitted 2026-06-26 · ❄️ cond-mat.mtrl-sci

Surrogate-Gated Generation and Foundation-Model Embeddings for Bayesian Materials Design

Sk Md Ahnaf Akif Alvi , Jan Janssen , Danny Perez , Douglas Allaire , Raymundo Arroyave This is my paper

Pith reviewed 2026-06-30 00:30 UTC · model grok-4.3

classification ❄️ cond-mat.mtrl-sci
keywords materials discoverygenerative modelsGaussian processsurrogate modelfoundation embeddingsBayesian designdiffusion priorscrystal structures
0
0 comments X

The pith

A Gaussian process surrogate on foundation embeddings gates generative proposals to match full-oracle results at one-fifth the cost.

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

The paper inserts a cheap probabilistic surrogate between a learned generative prior and an expensive property oracle in closed-loop materials discovery. The surrogate triages the generator's output so that only a budgeted number of candidates reach the oracle each cycle. Across multiple diffusion-based generators and targets like heat capacity and bulk modulus, the gated workflow matches or exceeds the performance of ungated fine-tuning. A sympathetic reader would care because oracle evaluations dominate discovery cost, and reliable triage could multiply the number of cycles possible within fixed resources.

Core claim

A Gaussian process acquisition gate placed between structure generation and the oracle in an RL-steered workflow matches or exceeds ungated fine-tuning while limiting oracle calls to a fixed budget per cycle. At a four-call budget the gate reaches within approximately 9 percent of exhaustive oracle performance using roughly one-fifth the calls, with the surrogate's ranking choice driving the gain over arbitrary selection.

What carries the argument

Gaussian process surrogate trained on ORB embeddings that performs ranking-based selection to decide which generated structures to evaluate with the oracle.

If this is right

  • Ranking-based selection from the surrogate outperforms arbitrary selection at the same budget.
  • The gate performs close to exhaustive oracle use at much lower cost across three distinct diffusion priors.
  • DFT validation confirms the learned oracle to within 2.5% and the surrogate ranking at Spearman rho of 0.94.
  • ORB embeddings with Gaussian process form the most reliable surrogate combination across mechanical, electronic, and vibrational properties.

Where Pith is reading between the lines

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

  • If the surrogate ranking holds under larger distribution shifts, the method could support longer RL-steered campaigns without retraining.
  • The open pipeline could be tested directly on additional generators or property oracles to measure generalization.
  • Extending the gate to multi-property oracles might allow simultaneous optimization of several targets at fixed budget.

Load-bearing premise

The Gaussian process surrogate maintains reliable ranking of structures sampled from the generative priors even when distribution shift occurs in the RL workflow.

What would settle it

Running the workflow with the surrogate gate and finding that the top structures selected do not achieve property values within 9% of those found by exhaustive oracle evaluation at the same total budget.

Figures

Figures reproduced from arXiv: 2606.28578 by Danny Perez, Douglas Allaire, Jan Janssen, Raymundo Arroyave, Sk Md Ahnaf Akif Alvi.

Figure 1
Figure 1. Figure 1: Modular pipeline for foundation-model-based [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The closed-loop pipeline. The diffusion gen [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Discovery versus oracle cost, per backbone: running best oracled property (seed-mean) against cumulative [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: GP CV5 RMSE (normalized property scale) per RL cycle, ACC runs only, seed-mean [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: MatterGen oracle-budget ablation as discovery versus cost: running best oracled property (seed-mean) [PITH_FULL_IMAGE:figures/full_fig_p010_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Crystal-system distribution among top-3 generated [PITH_FULL_IMAGE:figures/full_fig_p011_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Crystal-system distribution among the top-3 generated [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Per-backbone summary across the completed closed-loop runs (29 [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Top-20 generated structures per target (4 [PITH_FULL_IMAGE:figures/full_fig_p013_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: ORB-PU versus CGNF synthesizability scores [PITH_FULL_IMAGE:figures/full_fig_p013_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Per-(backbone, policy) synthesizability of the top generated structures ( [PITH_FULL_IMAGE:figures/full_fig_p014_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: DFT validation of the bulk-modulus discoveries. (a) Oracle fidelity: the eSEN oracle [PITH_FULL_IMAGE:figures/full_fig_p015_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Averaged R2 across all target properties at ntrain = 500, grouped by featurizer and surrogate model, for each of the three benchmark datasets. ORB embeddings rank first in most featurizer–surrogate cells; SOAP edges ORB on the dielectric dataset under GP and MTGP. Note the different y-axis scales, which reflect the different R2 ranges across datasets [PITH_FULL_IMAGE:figures/full_fig_p016_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Learning curves showing averaged R2 as a function of training set size for the GP, MTGP, and DGP surrogates on ORB embeddings, elastic tensor dataset. ORB + GP at ntrain = 100 approaches the performance of SOAP + GP at ntrain = 500 ( [PITH_FULL_IMAGE:figures/full_fig_p017_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Property difficulty matrices showing the best-achievable R [PITH_FULL_IMAGE:figures/full_fig_p018_15.png] view at source ↗
read the original abstract

Closed-loop materials discovery iterates between proposing candidate structures and evaluating their properties, and property evaluation dominates the cost. In the generative variant, a learned prior proposes candidate crystals and a property oracle scores them; we ask whether a cheap probabilistic surrogate can triage the generator's output, and what such a surrogate must do well. Across three architecturally distinct pretrained diffusion priors (MatterGen, CrystalFlow, ADiT) and two targets (room-temperature heat capacity and bulk modulus), we insert a Gaussian process acquisition gate between structure generation and the oracle in an RL-steered generative workflow. The gate matches or exceeds ungated fine-tuning of the generative model while capping oracle calls at a fixed per-cycle budget. Budget-matched ablations isolate the mechanism. At an identical four-call budget, ranking-based selection outperforms arbitrary selection, confirming that the gain comes from the surrogate's choice; the gate comes within $\sim$9\% of exhaustive oracle spending at roughly one-fifth of the calls. A density-functional-theory check of the bulk-modulus discoveries confirms the learned oracle to within 2.5\% on average and the surrogate's ranking of the generated structures at Spearman $\rho = 0.94$. A cross-factorial benchmark of surrogate performance spanning mechanical, electronic, and vibrational properties identifies pretrained ORB embeddings with a Gaussian process as the most reliable combination, which we adopt as the building blocks of the proposed workflow. The complete pipeline is released as open-source software.

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

0 major / 2 minor

Summary. The manuscript introduces a surrogate-gated workflow for closed-loop generative materials discovery. A Gaussian process surrogate operating on pretrained ORB embeddings is inserted as an acquisition gate between diffusion-based structure generators (MatterGen, CrystalFlow, ADiT) and an expensive property oracle within an RL-steered loop. The approach is tested on room-temperature heat capacity and bulk modulus, with budget-matched ablations, a cross-factorial surrogate benchmark, and DFT validation of discovered structures (2.5% average error, Spearman ρ=0.94). The gate is reported to match or exceed ungated fine-tuning while limiting oracle calls to a fixed per-cycle budget, and the full pipeline is released as open-source software.

Significance. If the empirical results hold, the work demonstrates a practical, budget-controlled mechanism for integrating cheap probabilistic surrogates into generative priors, reducing the dominant cost of oracle evaluations without sacrificing ranking quality. The open-source release, DFT cross-check, and identification of ORB+GP as the strongest embedding-surrogate pair across mechanical/electronic/vibrational properties provide concrete, reproducible building blocks for the field.

minor comments (2)
  1. [Abstract] Abstract: the phrase 'RL-steered generative workflow' is used without specifying the reinforcement-learning algorithm, reward formulation, or exact integration point of the gate; a brief clause would improve clarity for readers unfamiliar with the prior literature.
  2. [Abstract] The cross-factorial benchmark is described as selecting ORB+GP, but the abstract does not report the number of property classes, number of embedding models, or statistical significance of the ranking; adding these details would strengthen the claim.

Simulated Author's Rebuttal

0 responses · 0 unresolved

We thank the referee for their positive assessment of the work and the recommendation to accept.

Circularity Check

0 steps flagged

No significant circularity detected

full rationale

The paper's central claims rest on empirical DFT validations (2.5% oracle error, Spearman ρ=0.94 on generated structures), budget-matched ablations isolating ranking vs. arbitrary selection, and a cross-factorial benchmark identifying ORB+GP as optimal before adoption. These elements are externally falsifiable and do not reduce any prediction or gate performance to a fitted parameter or self-citation by construction. No self-definitional steps, fitted-input predictions, or load-bearing self-citations appear in the reported workflow; the derivation chain remains self-contained against the stated external checks and open-source release.

Axiom & Free-Parameter Ledger

2 free parameters · 1 axioms · 0 invented entities

The approach rests on standard ML assumptions about transferability of pretrained embeddings and suitability of GPs as surrogates; no new physical entities are postulated.

free parameters (2)
  • GP kernel hyperparameters
    Fitted to surrogate training data for each property and embedding combination.
  • acquisition function threshold or budget parameters
    Chosen to enforce fixed per-cycle oracle budget.
axioms (1)
  • domain assumption Pretrained ORB embeddings capture transferable structural features relevant to mechanical, electronic, and vibrational properties
    Adopted as the building blocks after the cross-factorial benchmark.

pith-pipeline@v0.9.1-grok · 5811 in / 1222 out tokens · 39913 ms · 2026-06-30T00:30:59.463698+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

53 extracted references · 33 canonical work pages · 1 internal anchor

  1. [1]

    Generative AI for crystal structures: a review.npj Computational Materials, 2025

    Pierre-Paul De Breuck, Hai-Chen Wang, Gian- Marco Rignanese, Silvana Botti, and Miguel A L Marques. Generative AI for crystal structures: a review.npj Computational Materials, 2025. doi: 10.1038/s41524-025-01881-2

    work page doi:10.1038/s41524-025-01881-2 2025

  2. [2]

    Crystal diffusion variational autoencoder for periodic material gen- eration

    Tian Xie, Xiang Fu, Octavian-Eugen Ganea, Regina Barzilay, and Tommi Jaakkola. Crystal diffusion variational autoencoder for periodic material gen- eration. InInternational Conference on Learning Representations, 2022. arXiv:2110.06197. 20

    work page arXiv 2022

  3. [3]

    Crystal struc- ture prediction by joint equivariant diffusion

    Rui Jiao, Wenbing Huang, Peijia Lin, Jiaqi Han, Pin Chen, Yutong Lu, and Yang Liu. Crystal struc- ture prediction by joint equivariant diffusion. InAd- vances in Neural Information Processing Systems,

  4. [4]

    A generative model for inor- ganic materials design.Nature, 639(8055):624–632,

    Claudio Zeni, Robert Pinsler, Daniel Z¨ ugner, An- drew Fowler, Matthew Horton, Xiang Fu, Zilong Wang, Aliaksandra Shysheya, Jonathan Crabb´ e, Shoko Ueda, et al. A generative model for inor- ganic materials design.Nature, 639(8055):624–632,

  5. [5]

    Benjamin Kurt Miller, Ricky T. Q. Chen, Anuroop Sriram, and Brandon M Wood. FlowMM: Generat- ing materials with Riemannian flow matching. InIn- ternational Conference on Machine Learning, 2024. arXiv:2406.04713

    work page arXiv 2024

  6. [6]

    CrystalFlow: a flow-based gen- erative model for crystalline materials.Nature Communications, 16(1):9267, 2025

    Xiaoshan Luo, Zhenyu Wang, Qingchang Wang, Xuechen Shao, Jian Lv, Lei Wang, Yanchao Wang, and Yanming Ma. CrystalFlow: a flow-based gen- erative model for crystalline materials.Nature Communications, 16(1):9267, 2025. doi: 10.1038/ s41467-025-64364-4

    2025

  7. [7]

    All-atom diffu- sion transformers: Unified generative modelling of molecules and materials.arXiv preprint arXiv:2503.03965, 2025

    Chaitanya K Joshi, Xiang Fu, Yi-Lun Liao, Vahe Gharakhanyan, Benjamin Kurt Miller, Anuroop Sriram, and Zachary W Ulissi. All-atom diffu- sion transformers: Unified generative modelling of molecules and materials.arXiv preprint arXiv:2503.03965, 2025

    work page arXiv 2025

  8. [8]

    Training Diffusion Models with Reinforcement Learning

    Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training dif- fusion models with reinforcement learning. In International Conference on Learning Represen- tations, volume 2024, pages 4965–4987, 2024. arXiv:2305.13301

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  9. [9]

    DPOK: Reinforcement learn- ing for fine-tuning text-to-image diffusion mod- els

    Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Mohammad Ghavamzadeh, Kangwook Lee, and Kimin Lee. DPOK: Reinforcement learn- ing for fine-tuning text-to-image diffusion mod- els. InAdvances in Neural Information Process- ing Systems, volume 36, pages 79858–79885, 2023. arXiv:2305.16381

    work page arXiv 2023

  10. [10]

    Accelerating inverse materials design us- ing generative diffusion models with reinforcement learning.arXiv preprint arXiv:2511.03112, 2025

    Junwu Chen, Jeff Guo, Edvin Fako, and Philippe Schwaller. Accelerating inverse materials design us- ing generative diffusion models with reinforcement learning.arXiv preprint arXiv:2511.03112, 2025

    work page arXiv 2025

  11. [11]

    CrystalFormer- RL: Reinforcement fine-tuning for materials design

    Zhendong Cao and Lei Wang. CrystalFormer- RL: Reinforcement fine-tuning for materials design. arXiv preprint arXiv:2504.02367, 2025

    work page arXiv 2025

  12. [12]

    Uehara, Y

    Masatoshi Uehara, Yulai Zhao, Tommaso Bian- calani, and Sergey Levine. Understanding rein- forcement learning-based fine-tuning of diffusion models: A tutorial and review.arXiv preprint arXiv:2407.13734, 2024

    work page arXiv 2024

  13. [13]

    Latent conserva- tive objective models for data-driven crystal struc- ture prediction.arXiv preprint arXiv:2310.10056, 2023

    Han Qi, Xinyang Geng, Stefano Rando, Iku Ohama, Aviral Kumar, and Sergey Levine. Latent conserva- tive objective models for data-driven crystal struc- ture prediction.arXiv preprint arXiv:2310.10056, 2023

    work page arXiv 2023

  14. [14]

    A mobile robotic chemist.Nature, 583(7815):237–241, 2020

    Benjamin Burger, Phillip M Maffettone, Vladimir V Gusev, Catherine M Aitchison, Yang Bai, Xiaoyan Wang, Xiaobo Li, Ben M Alston, Buyi Li, Rob Clowes, Nicola Rankin, Brandon Harris, Reiner Se- bastian Sprick, and Andrew I Cooper. A mobile robotic chemist.Nature, 583(7815):237–241, 2020. doi: 10.1038/s41586-020-2442-2

    work page doi:10.1038/s41586-020-2442-2 2020

  15. [15]

    A tutorial on Bayesian optimization

    Peter I Frazier. A tutorial on Bayesian optimization. Recent Advances in Optimization and Modeling of Contemporary Problems, pages 255–278, 2018. doi: 10.1287/educ.2018.0188. arXiv:1807.02811

    work page doi:10.1287/educ.2018.0188 2018

  16. [16]

    Spectral Deferred Correction Methods for Ordinary Differential Equations

    Donald R Jones, Matthias Schonlau, and William J Welch. Efficient global optimization of expen- sive black-box functions.Journal of Global Op- timization, 13(4):455–492, 1998. doi: 10.1023/A: 1008306431147

    work page doi:10.1023/a: 1998

  17. [17]

    MACE: Higher order equivariant message passing neural networks for fast and accurate force fields

    Ilyes Batatia, D´ avid P´ eter Kov´ acs, Gregor N C Simm, Christoph Ortner, and G´ abor Cs´ anyi. MACE: Higher order equivariant message passing neural networks for fast and accurate force fields. InAdvances in Neural Information Processing Sys- tems, volume 35, pages 11423–11436, 2022. doi: 10.52202/068431-0830

    work page doi:10.52202/068431-0830 2022

  18. [18]

    A foundation model for atomistic materials chemistry.The Journal of chem- ical physics, 163(18), 2025

    Ilyes Batatia, Philipp Benner, Yuan Chiang, Alin M Elena, D´ avid P Kov´ acs, Janosh Riebesell, Xavier R Advincula, Mark Asta, Matthew Avaylon, William J Baldwin, et al. A foundation model for atomistic materials chemistry.The Journal of chem- ical physics, 163(18), 2025. doi: 10.1063/5.0297006

    work page doi:10.1063/5.0297006 2025

  19. [19]

    Orb: A fast, scalable neural network potential.arXiv preprint arXiv:2410.22570, 2024

    Mark Neumann, James Gin, Benjamin Rhodes, Steven Bennett, Zhiyi Li, Hitarth Choubisa, Arthur Hussey, and Jonathan Godwin. Orb: A fast, scalable neural network potential.arXiv preprint arXiv:2410.22570, 2024

    work page arXiv 2024

  20. [20]

    Learning smooth and expres- sive interatomic potentials for physical property pre- diction.arXiv preprint arXiv:2502.12147, 2025

    Xiang Fu, Brandon M Wood, Luis Barroso-Luque, Daniel S Levine, Meng Gao, Misko Dzamba, and C Lawrence Zitnick. Learning smooth and expres- sive interatomic potentials for physical property pre- diction.arXiv preprint arXiv:2502.12147, 2025

    work page arXiv 2025

  21. [21]

    MIT Press, Cambridge, MA, 2006

    Carl Edward Rasmussen and Christopher K I Williams.Gaussian Processes for Machine Learn- ing. MIT Press, Cambridge, MA, 2006. doi: 10.7551/mitpress/3206.001.0001. 21

    work page doi:10.7551/mitpress/3206.001.0001 2006

  22. [22]

    Benchmarking materials property prediction methods: the Matbench test set and Automatminer reference algorithm.npj Com- putational Materials, 6(1):138, 2020

    Alexander Dunn, Qi Wang, Alex Ganose, Daniel Dopp, and Anubhav Jain. Benchmarking materials property prediction methods: the Matbench test set and Automatminer reference algorithm.npj Com- putational Materials, 6(1):138, 2020. doi: 10.1038/ s41524-020-00406-3

    2020

  23. [23]

    Springer, New York, 2nd edition, 2002

    Ian T Jolliffe.Principal Component Analysis. Springer, New York, 2nd edition, 2002

    2002

  24. [24]

    Matminer: An open source toolkit for materi- als data mining.Computational Materials Science, 152:60–69, 2018

    Logan Ward, Alexander Dunn, Alireza Faghaninia, Nils E R Zimmermann, Saurabh Bajaj, Qi Wang, Joseph H Montoya, Jiming Chen, Kyle Bystrom, Mark Dylla, Kyle Chard, Mark Asta, Kristin A Persson, G Jeffrey Snyder, Ian Foster, and Anubhav Jain. Matminer: An open source toolkit for materi- als data mining.Computational Materials Science, 152:60–69, 2018. doi: 1...

    work page doi:10.1016/j.commatsci.2018.05 2018

  25. [25]

    Guido Petretto, Shyam Dwaraknath, Henrique P. C. Miranda, Donald Winston, Matteo Giantomassi, Michiel J. van Setten, Xavier Gonze, Kristin A. Persson, Geoffroy Hautier, and Gian-Marco Rig- nanese. High-throughput density-functional pertur- bation theory phonons for inorganic materials.Sci- entific Data, 5:180065, 2018. doi: 10.1038/sdata. 2018.65

    work page doi:10.1038/sdata 2018

  26. [26]

    Com- mentary: The materials project: A materials genome approach to accelerating materials inno- vation.APL Materials, 1(1):011002, 2013

    Anubhav Jain, Shyue Ping Ong, Geoffroy Hautier, Wei Chen, William Davidson Richards, Stephen Dacek, Shreyas Cholia, Dan Gunter, David Skin- ner, Gerbrand Ceder, and Kristin A Persson. Com- mentary: The Materials Project: A materials genome approach to accelerating materials inno- vation.APL Materials, 1(1):011002, 2013. doi: 10.1063/1.4812323

    work page doi:10.1063/1.4812323 2013

  27. [27]

    Charting the complete elastic properties of inorganic crystalline compounds.Scientific Data, 2(1):150009, 2015

    Maarten de Jong, Wei Chen, Thomas Angsten, Anubhav Jain, Randy Notestine, Anthony Gamst, Marcel Sluiter, Chaitanya Krishna Ande, Sybrand van der Zwaag, Jose J Plata, Cormac Toher, Stefano Curtarolo, Gerbrand Ceder, Kristin A Persson, and Mark Asta. Charting the complete elastic properties of inorganic crystalline compounds.Scientific Data, 2(1):150009, 20...

    work page doi:10.1038/sdata.2015.9 2015

  28. [28]

    High-throughput screening of inorganic com- pounds for the discovery of novel dielectric and op- tical materials.Scientific Data, 4(1):160134, 2017

    Ioannis Petousis, David Mrdjenovich, Eric Ballouz, Miao Liu, Donald Winston, Wei Chen, Tanja Graf, Thomas D Schladt, Kristin A Persson, and Fritz B Prinz. High-throughput screening of inorganic com- pounds for the discovery of novel dielectric and op- tical materials.Scientific Data, 4(1):160134, 2017. doi: 10.1038/sdata.2016.134

    work page doi:10.1038/sdata.2016.134 2017

  29. [29]

    On representing chemical environments.Phys- ical Review B—Condensed Matter and Materi- als Physics, 87(18):184115, 2013

    Albert P Bart´ ok, Risi Kondor, and G´ abor Cs´ anyi. On representing chemical environments.Phys- ical Review B—Condensed Matter and Materi- als Physics, 87(18):184115, 2013. doi: 10.1103/ PhysRevB.87.184115

    2013

  30. [30]

    DScribe: Library of descriptors for machine learning in materials science.Computer Physics Communi- cations, 247:106949, 2020

    Lauri Himanen, Marc O J J¨ ager, Eiaki V Morooka, Filippo Federici Canova, Yashasvi S Ranawat, David Z Gao, Patrick Rinke, and Adam S Foster. DScribe: Library of descriptors for machine learning in materials science.Computer Physics Communi- cations, 247:106949, 2020. doi: 10.1016/j.cpc.2019. 106949

    work page doi:10.1016/j.cpc.2019 2020

  31. [31]

    Bartel, and Gerbrand Ceder

    Bowen Deng, Peichen Zhong, KyuJung Jun, Janosh Riebesell, Kevin Han, Christopher J Bartel, and Gerbrand Ceder. CHGNet as a pretrained uni- versal neural network potential for charge-informed atomistic modelling.Nature Machine Intelligence, 5: 1031–1041, 2023. doi: 10.1038/s42256-023-00716-3

    work page doi:10.1038/s42256-023-00716-3 2023

  32. [32]

    BoTorch: A frame- work for efficient monte-carlo Bayesian optimiza- tion

    Maximilian Balandat, Brian Karrer, Daniel R Jiang, Samuel Daulton, Benjamin Letham, Andrew Gor- don Wilson, and Eytan Bakshy. BoTorch: A frame- work for efficient monte-carlo Bayesian optimiza- tion. InAdvances in Neural Information Processing Systems, volume 33, pages 21524–21538, 2020

    2020

  33. [33]

    GPyTorch: Blackbox matrix-matrix Gaussian process inference with GPU acceleration

    Jacob R Gardner, Geoff Pleiss, Kilian Q Wein- berger, David Bindel, and Andrew Gordon Wil- son. GPyTorch: Blackbox matrix-matrix Gaussian process inference with GPU acceleration. InAd- vances in Neural Information Processing Systems, volume 31, 2018

    2018

  34. [34]

    Multi-task Gaussian pro- cess prediction

    Edwin V Bonilla, Kian Ming Adam Chai, and Christopher K I Williams. Multi-task Gaussian pro- cess prediction. InAdvances in Neural Information Processing Systems, volume 20, 2007

    2007

  35. [35]

    Deep Gaussian processes

    Andreas Damianou and Neil D Lawrence. Deep Gaussian processes. InProceedings of the Sixteenth International Conference on Artificial Intelligence and Statistics, pages 207–215, 2013

    2013

  36. [36]

    Deep neural networks as point es- timates for deep Gaussian processes

    Vincent Dutordoir, James Hensman, Mark van der Wilk, Carl Henrik Ek, Zoubin Ghahramani, and Nicolas Durrande. Deep neural networks as point es- timates for deep Gaussian processes. InAdvances in Neural Information Processing Systems, volume 34, 2021

    2021

  37. [37]

    First principles phonon calculations in materials science.Scripta Materialia, 108:1–5, 2015

    Atsushi Togo and Isao Tanaka. First principles phonon calculations in materials science.Scripta Materialia, 108:1–5, 2015

    2015

  38. [38]

    Finite elastic strain of cubic crystals

    Francis Birch. Finite elastic strain of cubic crystals. Physical Review, 71(11):809–824, 1947

    1947

  39. [39]

    MADE: Benchmark environments for closed-loop materials discovery

    Shreshth A Malik, Tiarnan Doherty, Panagiotis Tigas, Muhammed Razzak, Stephen J Roberts, Aron Walsh, and Yarin Gal. MADE: Benchmark environments for closed-loop materials discovery. arXiv preprint arXiv:2601.20996, 2026. 22

    work page arXiv 2026

  40. [40]

    LeMat-GenBench: A unified evaluation framework for crystal generative models.arXiv preprint arXiv:2512.04562, 2025

    Siddharth Betala, Samuel P Gleason, Ali Ram- laoui, Andy Xu, Georgia Channing, Daniel Levy, Cl´ ementine Fourrier, Nikita Kazeev, Chaitanya K Joshi, S´ ekou-Oumar Kaba, F´ elix Therrien, Alex Hernandez-Garcia, Roc´ ıo Mercado, N M Anoop Kr- ishnan, and Alexandre Duval. LeMat-GenBench: A unified evaluation framework for crystal generative models.arXiv prep...

    work page arXiv 2025

  41. [41]

    Spglib: a software library for crystal symmetry search.arXiv preprint arXiv:1808.01590, 2018

    Atsushi Togo and Isao Tanaka. Spglib: a software library for crystal symmetry search.arXiv preprint arXiv:1808.01590, 2018

    work page arXiv 2018

  42. [42]

    Efficient iter- ative schemes for ab initio total-energy calculations using a plane-wave basis set.Physical Review B, 54 (16):11169–11186, 1996

    Georg Kresse and J¨ urgen Furthm¨ uller. Efficient iter- ative schemes for ab initio total-energy calculations using a plane-wave basis set.Physical Review B, 54 (16):11169–11186, 1996

    1996

  43. [43]

    Python materials genomics (pymatgen): A robust, open- source python library for materials analysis.Com- putational Materials Science, 68:314–319, 2013

    Shyue Ping Ong, William Davidson Richards, Anubhav Jain, Geoffroy Hautier, Michael Kocher, Shreyas Cholia, Dan Gunter, Vincent L Chevrier, Kristin A Persson, and Gerbrand Ceder. Python materials genomics (pymatgen): A robust, open- source python library for materials analysis.Com- putational Materials Science, 68:314–319, 2013

    2013

  44. [44]

    Perdew, Kieron Burke, and Matthias Ernz- erhof

    John P. Perdew, Kieron Burke, and Matthias Ernz- erhof. Generalized gradient approximation made simple.Physical Review Letters, 77(18):3865–3868, 1996

    1996

  45. [45]

    Bl¨ ochl

    Peter E. Bl¨ ochl. Projector augmented-wave method. Physical Review B, 50(24):17953–17979, 1994

    1994

  46. [46]

    From ultrasoft pseudopotentials to the projector augmented-wave method.Physical Review B, 59(3):1758–1775, 1999

    Georg Kresse and Daniel Joubert. From ultrasoft pseudopotentials to the projector augmented-wave method.Physical Review B, 59(3):1758–1775, 1999

    1999

  47. [47]

    Gregoire, and Yousung Jung

    Jidon Jang, Juhwan Noh, Lan Zhou, Geun Ho Gu, John M. Gregoire, and Yousung Jung. Syn- thesizability of materials stoichiometry using semi- supervised learning.Matter, 7(6):2294–2312, 2024. doi: 10.1016/j.matt.2024.05.002

    work page doi:10.1016/j.matt.2024.05.002 2024

  48. [48]

    A general-purpose machine learning framework for predicting properties of inorganic materials

    Logan Ward, Ankit Agrawal, Alok Choudhary, and Christopher Wolverton. A general-purpose machine learning framework for predicting properties of inor- ganic materials.npj Computational Materials, 2(1): 16028, 2016. doi: 10.1038/npjcompumats.2016.28

    work page doi:10.1038/npjcompumats.2016.28 2016

  49. [49]

    A bag- ging svm to learn from positive and unlabeled exam- ples.Pattern Recognition Letters, 37:201–209, 2014

    Fantine Mordelet and Jean-Philippe Vert. A bag- ging svm to learn from positive and unlabeled exam- ples.Pattern Recognition Letters, 37:201–209, 2014. doi: 10.1016/j.patrec.2013.06.010

    work page doi:10.1016/j.patrec.2013.06.010 2014

  50. [50]

    In: Proceedings of the 25th ACM 3 For instance, see the technical report at https://www.anthropic.com/ claude-3-model-card

    Takuya Akiba, Shotaro Sano, Toshihiko Yanase, Takeru Ohta, and Masanori Koyama. Optuna: A next-generation hyperparameter optimization framework. InProceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 2623–2631, 2019. doi: 10.1145/3292500.3330701

    work page doi:10.1145/3292500.3330701 2019

  51. [51]

    Predicting materi- als properties with little data using shotgun trans- fer learning.ACS Central Science, 5(10):1717–1730,

    Hironao Yamada, Chang Liu, Stephen Wu, Yukinori Koyama, Shenghong Ju, Junichiro Shiomi, Junko Morikawa, and Ryo Yoshida. Predicting materi- als properties with little data using shotgun trans- fer learning.ACS Central Science, 5(10):1717–1730,

  52. [52]

    Supplementary Fig. Sn

    doi: 10.1021/acscentsci.9b00804. 23 Supporting Information for Surrogate-Gated Diffusion and Foundation-Model Embeddings for Bayesian Materials Design Sk Md Ahnaf Akif Alvi, Jan Janssen, Danny Perez, Douglas Allaire, Raymundo Arr´ oyave This Supplementary Information collects the closed-loop per-cycle views and the static- benchmark deep-dive figures and ...

    work page doi:10.1021/acscentsci.9b00804

  53. [53]

    upper bound

    Each bar shows the best R 2 (a) or Spearmanρ(b) across all surrogate–PCA combinations, with the winning surrogate name overlaid. ORB achieves the highest R 2 on all three datasets, and its Spearman correlations stay above 0.7 across datasets. 7 Figure S13: Best averaged R 2 per surrogate across the three datasets (ORB descriptor,n train = 500, best PCA pe...

    2000