acceptodds
Under review as a conference paper at ICLR 2027

Multi-property Feynman-Kac Steering of Crystal Diffusion Models

Abstract

Pretrained diffusion models provide powerful priors for generating new materials, including inorganic crystals, but generating useful materials requires satisfying specific chemical and physical properties, from conductivity to band gaps. Existing approaches either retrain models for new target properties or steer only the atom coordinates and lattice during generation, leaving composition unchanged despite its fundamental role in determining physical properties. We present FK-ReMask, an inference-time, gradient-free method that steers both the continuous and discrete channels of a pretrained crystal diffusion model. During generation, FK-ReMask maintains multiple denoising trajectories and replicates promising candidates using Feynman–Kac particle selection. However, selection alone is insufficient: in the atom-type channel, committed elements cannot be revised, so cloning propagates existing compositions and limits the population's ability to explore new ones. We address this with an atom re-masking mechanism, which reopens committed sites during the generation process and allows alternative compositions to emerge. Across four electronic, photovoltaic, and mechanical objectives, our method improves the yield of SUN (stable, unique, novel) and property-on-target materials. We further demonstrate joint steering of three photovoltaic properties, showing that multiple physical objectives can be specified entirely at inference time without retraining.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
inference-time steering, diffusion models, multi-objective optimization, generative materials design

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_NoWwOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper studies the problem of inverse design of crystals with target properties, with a technical focus on training-free steering of generative models. The paper argues that existing methods that can steer the inference of a generative model only work on the continuous channels of crystal structures (coordinates and lattices), but do not work on the discrete channels (composition, or atomic elements). The paper proposes FK-ReMask that can steer both continuous and discrete channels during the generation of crystal structures without re-training the model.

Critical strengths
  1. The paper tackles a relevant research problem in the field of crystal material design with a straightforward method.
Critical weaknesses
  1. The technical novelty of the proposed method can be limited. On paper, most of the technical designs in the proposed method are derived from existing techniques from either the crystal material design field or other fields (most of which are discussed in the paper's related work, so kudos to the authors for that).
  2. The proposed method is built on top of MatterGen, and the proposed method is only compared to the vanilla MatterGen, along with a few ablations of it. It could provide more empirical evidence if more conditional crystal structure generation methods could be compared.
Questions for authors
  1. Could the authors reiterate the technical contribution of their proposed method?
  2. Could the authors debate the rationale for only comparing with the vanilla MatterGen and its variations in the experiments, or even better, provide experimental comparison with more methods?
3 comments

 

We thank the reviewer for the feedback and for appreciating our effort to position the proposed method in the context of existing work.

 

1. Our technical contribution:

 

  • Joint steering of discrete and continuous channels. FK-ReMask is the first inference-time steering method for crystal diffusion models that jointly steers discrete (atom types) and continuous (atomic positions and lattice) channels. It navigates chemical space using channel-specific perturbations that enable exploration while remaining close to the pretrained model's learned distribution: continuous channels retain the default diffusion noise schedule, while the discrete channel employs special re-masking (ReMDM) that preserves the pretrained masking schedule at every step.

 

  • Repurposing re-masking for population diversity. ReMDM was originally designed for error correction within a single trajectory. We repurpose it to preserve compositional diversity across multiple trajectories under FK selection, so clones can explore alternative compositions.

 

2. Following the reviewer's suggestion, we extend the comparison in three directions:

 

  • A second base model with a different conditioning mechanism. ChargeDIFF (Park et al., 2026) jointly generates atomic structures and electron charge densities. Unlike MatterGen's CFG-based conditioning, ChargeDIFF's band-gap-conditional model ("Conditional") directly incorporates the target into the denoiser, without classifier-free guidance.

     

    Applying FK-ReMask (k=8) to unconditional ChargeDIFF achieves a 9.6% SUN & on-target yield, compared with 3.1% for Best-of-N and 2.4% for Conditional, but it is below Conditional + Best-of-N (13.2%).

     

    FK-ReMask can also be applied on top of Conditional. Combined with a novelty reward, it achieves a SUN & on-target yield of 14.8%, outperforming Conditional + Best-of-N (13.2%) and achieving the highest yield among the evaluated methods. The latter combination comes at the cost of reduced validity (64.8%). Adding a validity term restores validity to 84.1%, the highest in the table, while keeping SUN & on-target at 14.7% (Appendix F, Table A6).

methodvaliduniquenovelstableSUNon-targetstable & on-targetSUN & on-target
ChargeDIFF83.9 ± 1.099.2 ± 0.164.1 ± 1.351.8 ± 0.721.0 ± 0.50.9 ± 0.40.4 ± 0.10.1 ± 0.1
+ Best-of-N83.7 ± 0.898.0 ± 0.269.6 ± 0.951.7 ± 0.825.3 ± 0.210.7 ± 0.75.7 ± 0.43.1 ± 0.5
+ Best-of-N + novelty81.6 ± 0.598.0 ± 0.178.6 ± 0.845.8 ± 0.727.5 ± 0.49.4 ± 0.94.8 ± 0.22.8 ± 0.5
+ Conditional82.2 ± 1.097.6 ± 0.769.9 ± 0.852.9 ± 1.826.3 ± 2.39.5 ± 1.36.6 ± 1.32.4 ± 1.0
+ Conditional + Best-of-N83.3 ± 0.991.1 ± 0.363.5 ± 1.363.3 ± 0.528.1 ± 1.353.4 ± 0.534.8 ± 0.613.2 ± 0.7
+ Conditional + Best-of-N + novelty74.7 ± 1.395.6 ± 0.383.6 ± 0.545.8 ± 0.830.3 ± 0.540.2 ± 0.721.6 ± 0.910.9 ± 0.6
+ FK-ReMask79.4 ± 0.189.1 ± 0.955.4 ± 1.165.7 ± 0.526.0 ± 1.135.7 ± 0.524.9 ± 1.19.6 ± 0.6
+ FK-ReMask + novelty72.3 ± 1.192.5 ± 0.775.1 ± 0.248.4 ± 0.626.5 ± 0.430.1 ± 1.015.4 ± 0.96.8 ± 0.5
+ Conditional + FK-ReMask81.1 ± 0.677.3 ± 0.547.7 ± 1.179.6 ± 1.529.6 ± 0.155.3 ± 0.946.7 ± 1.312.4 ± 0.2
+ Conditional + FK-ReMask + novelty64.8 ± 0.389.2 ± 0.576.2 ± 0.955.8 ± 0.834.0 ± 1.047.5 ± 0.830.2 ± 0.514.8 ± 0.3
+ Conditional + FK-ReMask + novelty + validity84.1 ± 0.386.5 ± 1.072.1 ± 0.861.0 ± 0.633.6 ± 0.348.0 ± 1.232.8 ± 0.714.7 ± 0.4

 

  • Additional training-free steering method. We compared FK-ReMask with Universal Guidance (UG). UG steers only atom positions and the lattice, whereas FK-ReMask also steers atom types, enabling control over properties of the entire crystal.

     

    For band gap and bulk modulus tasks, FK-ReMask achieves the highest on-target, stable & on-target, and SUN & on-target yields (single seed, 512 crystals each).

     

    We additionally evaluate a task from the UG paper: generating Li-Co-O crystals with a mean Co-O coordination number of 4. All methods use MatterGen conditioned on Li-Co-O with CFG, which encourages but does not enforce the presence of these elements. Consequently, 80% of vanilla MatterGen’s generated crystals contain no cobalt, making the target inaccessible through atomic-position steering alone. UG does not overcome this limitation (only 19% of its crystals contain cobalt), whereas FK-ReMask increases this fraction to 75% and achieves the highest yields on all three metrics.

ValidUniqueNovelStableSUNOn targetStable & on targetSUN & on target
Band gap, 2.4 ± 0.1 eV
MatterGen86.3 ± 1.599.6 ± 0.359.2 ± 2.275.0 ± 1.937.5 ± 2.10.4 ± 0.30.4 ± 0.30.2 ± 0.2
+ Universal Guidance86.1 ± 1.599.8 ± 0.259.6 ± 2.276.8 ± 1.939.6 ± 2.22.7 ± 0.72.1 ± 0.61.2 ± 0.5
+ FK-ReMask80.9 ± 1.796.1 ± 0.968.2 ± 2.182.6 ± 1.751.6 ± 2.264.5 ± 2.151.6 ± 2.232.6 ± 2.1
Bulk modulus, 150 ± 15 GPa
MatterGen86.3 ± 1.599.6 ± 0.359.2 ± 2.275.0 ± 1.937.5 ± 2.16.4 ± 1.14.1 ± 0.91.8 ± 0.6
+ Universal Guidance86.5 ± 1.599.4 ± 0.358.4 ± 2.277.7 ± 1.840.0 ± 2.26.1 ± 1.15.5 ± 1.02.0 ± 0.6
+ FK-ReMask76.0 ± 1.996.9 ± 0.852.7 ± 2.276.8 ± 1.933.4 ± 2.196.1 ± 0.974.0 ± 1.932.4 ± 2.1
Mean Co–O coordination, Li–Co–O, 4 ± 0.5
MatterGen, conditioned on Li–Co–O48.8 ± 2.260.9 ± 2.278.9 ± 1.837.7 ± 2.19.8 ± 1.35.1 ± 1.02.9 ± 0.82.0 ± 0.6
+ Universal Guidance42.0 ± 2.265.4 ± 2.183.4 ± 1.629.1 ± 2.010.9 ± 1.44.5 ± 0.90.8 ± 0.40.8 ± 0.4
+ FK-ReMask74.4 ± 1.960.7 ± 2.281.4 ± 1.738.9 ± 2.213.3 ± 1.537.1 ± 2.19.8 ± 1.35.7 ± 1.0

 

  • FK-ReMask as a complementary tool for conditional models. We applied FK-ReMask on top of MatterGen with classifier-free guidance (CFG), whose property adapter is fine-tuned on labelled data. On bandgap, and bulk modulus, it improves the SUN & on-target rates, in comparison to CFG and CFG + Best-of-N (Appendix G).
propertymethodvaliduniquenovelstableSUNon-targetstable & on-targetSUN & on-target
Bandgap (E_g)CFG86.3 ± 0.299.0 ± 0.269.0 ± 1.666.9 ± 1.237.7 ± 0.65.0 ± 0.33.3 ± 0.11.5 ± 0.3
CFG + Best-of-N86.0 ± 1.197.7 ± 0.268.9 ± 1.568.9 ± 1.039.5 ± 0.436.5 ± 0.925.2 ± 0.913.5 ± 0.4
CFG + FK-ReMask88.2 ± 1.590.7 ± 0.262.1 ± 1.086.1 ± 0.9*47.3 ± 0.5*63.3 ± 0.4*55.7 ± 0.8*29.9 ± 0.3*
Bulk Modulus (K)CFG81.2 ± 1.499.5 ± 0.157.1 ± 0.866.1 ± 1.327.9 ± 1.153.3 ± 0.336.8 ± 0.515.8 ± 0.2
CFG + Best-of-N78.7 ± 0.698.8 ± 0.255.2 ± 0.165.9 ± 1.327.3 ± 0.799.7 ± 0.165.7 ± 1.427.2 ± 0.8
CFG + FK-ReMask80.8 ± 0.595.3 ± 0.552.7 ± 1.576.4 ± 0.7*33.7 ± 1.499.7 ± 0.276.2 ± 0.6*33.5 ± 1.5*
Reviewer_3HMXOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

FK-ReMask is an inference-time, gradient-free method for steering a frozen crystal diffusion model (MatterGen) toward target material properties. It maintains a population of k denoising trajectories, periodically resamples them by a reward from off-the-shelf property predictors (Feynman–Kac / SMC selection), and adds an atom re-masking mechanism (the ReMDM operator) that reopens already-committed atom sites so cloned particles can explore new chemical compositions instead of collapsing onto one. Across four single properties and a joint three-property photovoltaic task, it reports large increases in the fraction of generated crystals that are simultaneously stable/unique/novel (SUN) and on-target.

Critical strengths
  • clean, model-agnostic, genuinely useful recipe. training-free retargeting of a frozen generator across both geometry and composition plugs into any absorbing-state crystal diffusion model with any off-the-shelf predictor. the re-masking-across-trajectories idea is well motivated and the collapse mechanism (§4.2) is crisply argued.
  • the re-masking kernel is mathematically sound. independent symbolic check: eq.1 coefficients are valid probabilities summing to 1, the committed-fraction identity holds, and the eq.2 cap tightly keeps the mask coefficient non-negative. the core math is not the weak point.
Critical weaknesses
  • reward = evaluator, with no independent verification. the fk reward and the on-target success metric both come from the same off-the-shelf alignn models. so the big "on-target" numbers (e.g. slme 3.1->79.1, bulk modulus 4.9->95.9) are "predictor-on-target", not "property-on-target". with no DFT and no held-out predictor, the central claim (real property control) is unproven. this one issue is decisive.

  • selective / incomplete reporting around the win: -- confidence intervals only on the two winning columns (n=512, binomial SE ~2.2pp makes several "wins" ties), -- validity/novelty/SUN regressions on bulk modulus never acknowledged in the prose, -- CFG baseline missing on 2 of 4 properties (so the biggest margins sit next to the weakest baseline), -- FK gets ~8x the compute CFG does with no NFE (number of function evaluations) table, -- and code is withheld. the gains are dressed up better than the evidence supports.

Questions for authors

can you address weaknesses from above?

  • if you steer with predictor A and evaluate on-target with an independent predictor B (different architecture/checkpoint), how much of the on-target gain survives?
  • DFT single-point+relax a random 30-50 of your "SUN & on-target" hits (and the baselines'): what is the DFT-confirmed rate with cis?
  • with Confidence Intervals on every column and ≥3 seeds, which of the SUN / novel / validity comparisons are still significant?
  • does τ->0 inflate surrogate-on-target while held-out on-target saturates (ie reward-hacking signature)?
  • head-to-head vs Hasan's resampling-suspension and vs t2c-fk / universal-guidance where they apply.
4 comments

We thank the reviewer for their detailed and constructive feedback. Their comments motivated additional experiments to provide a more comprehensive evaluation of FK-ReMask. We are encouraged that you find the approach to be clean and genuinely useful, the idea to be well motivated, and the math to be sound.

 

1) Selective/incomplete reporting around the win.

 

  • Code: We uploaded the code in a zip file.

 

  • Confidence Intervals and seeds: We ran the experiments with three random seeds. Tables 2 and 3 now report the mean over seeds ± its standard error. We added a star (*) to mark a result where FK-ReMask is significantly better than every other method, using a paired t-test (p < 0.05). We put the requested 95% confidence intervals in Appendix J.

 

  • CFG baseline missing on 2 of 4 properties: MatterGen did not publish a CFG adapter for SLME and hole effective mass.

 

  • FK vs. CFG compute. We added Table A11 in Appendix I, reporting the number of function evaluations (NFE) and total predictor calls required to generate a single crystal for each method in the single-objective bandgap task. For a fair comparison, we also added CFG + Best-of-N baselines with NFE matching or exceeding that of FK-ReMask. Both approaches therefore operate under the same NFE budget and select the best candidate from their respective k relaxed structures, differing only in the total number of predictor calls.

     

    For the comparison, we generated 3 x 512 crystals per method and averaged the results. FK-ReMask achieves higher SUN & on-target rates with only 16,000 NFE per crystal, compared to 16,000 and 32,000 NFE for CFG + Best-of-N (k=4) and CFG + Best-of-N (k=8), respectively:

methodNFEpredictor callsvaliduniquenovelstableSUNon-targetstable & on-targetSUN & on-target
MatterGen2,000086.0 ± 0.899.7 ± 0.159.7 ± 0.673.3 ± 1.437.0 ± 1.71.2 ± 0.40.8 ± 0.30.5 ± 0.2
CFG4,000086.3 ± 0.299.0 ± 0.269.0 ± 1.666.9 ± 1.237.7 ± 0.65.0 ± 0.33.3 ± 0.11.5 ± 0.3
Best-of-N (k=8)16,000882.4 ± 0.3100.0 ± 0.070.4 ± 1.469.2 ± 2.041.9 ± 0.98.7 ± 0.66.1 ± 0.44.2 ± 0.2
CFG + Best-of-N (k=4)16,000486.3 ± 1.498.8 ± 0.271.3 ± 1.068.4 ± 1.241.1 ± 0.621.1 ± 0.413.9 ± 0.36.8 ± 0.2
CFG + Best-of-N (k=8)32,000886.0 ± 1.197.7 ± 0.268.9 ± 1.568.9 ± 1.039.5 ± 0.436.5 ± 0.925.2 ± 0.913.5 ± 0.4
FK-ReMask (k=8)16,00028881.2 ± 0.396.5 ± 0.570.1 ± 1.283.8 ± 0.654.8 ± 1.763.9 ± 0.553.6 ± 1.135.4 ± 1.7

For the bulk modulus task, which also has a CFG baseline, we added CFG + Best-of-N (k=8) as a stronger baseline. FK-ReMask also achieves higher SUN & on-target rates in this comparison (Table 2).

  • Validity/novelty/SUN regressions on bulk modulus: We added this issue in the text (section 6.1), and also show how to address this (details in the next bullet).

     

  • With confidence intervals on every column and ≥3 seeds, which of the SUN / novel / validity comparisons are still significant?

     

    We ran the four single-objective tasks (Table 2) and the joint three-objective task (Table 3) with three random seeds and a significance test as described above. We call FK-ReMask competitive when it is not significantly worse than any baseline.

     

    Most importantly, FK-ReMask is significantly better on the main metric, SUN & on-target, in four of the five tasks, and competitive for bulk modulus. In more detail:

     

    • Targeting metrics: significantly better on all three targeting metrics (on-target, stable & on-target, SUN & on-target) for band gap, SLME, and hole mass, and on all six targeting metrics of the three-objective solar task. For bulk modulus, significantly better on stable & on-target, competitive on SUN & on-target, and significantly worse on on-target, where CFG + Best-of-N reaches 99.7% against our 97.0%.

     

    • SUN: significantly better for band gap, SLME, and hole mass, competitive for the three-objective task, and significantly worse for bulk modulus.

     

    • Validity: competitive for SLME, hole mass, and the three-objective task, and significantly worse for band gap and bulk modulus.

     

    • Novelty: competitive in all five tasks. For bulk modulus, FK-ReMask has the lowest mean, but the difference is not significant.

     

    As written above, in some cases steering with FK-ReMask degrades metrics like validity and novelty. We show that this can often be remedied directly by adding appropriate reward terms (Appendix E). For example, on bulk modulus, adding validity and novelty terms makes FK-ReMask restore them and also get the highest SUN & on-target rate:

methodvaliduniquenovelstableSUNon-targetstable & on-targetSUN & on-target
MatterGen86.0 ± 0.899.7 ± 0.159.7 ± 0.673.3 ± 1.437.0 ± 1.76.0 ± 0.23.5 ± 0.41.3 ± 0.2
+ Best-of-N80.7 ± 0.799.7 ± 0.159.8 ± 0.962.5 ± 1.027.7 ± 0.834.3 ± 1.620.8 ± 1.58.6 ± 1.1
+ Best-of-N + validity + novelty85.4 ± 1.199.9 ± 0.161.8 ± 1.162.2 ± 0.829.0 ± 1.033.8 ± 1.620.5 ± 1.48.9 ± 0.9
+ CFG81.2 ± 1.499.5 ± 0.157.1 ± 0.866.1 ± 1.327.9 ± 1.153.3 ± 0.336.8 ± 0.515.8 ± 0.2
+ CFG + Best-of-N78.7 ± 0.698.8 ± 0.255.2 ± 0.165.9 ± 1.327.3 ± 0.799.7 ± 0.165.7 ± 1.427.2 ± 0.8
+ CFG + Best-of-N + validity + novelty97.5 ± 0.299.2 ± 0.271.9 ± 0.760.4 ± 1.235.5 ± 0.899.0 ± 0.559.8 ± 1.635.0 ± 1.1
+ FK-ReMask76.6 ± 0.896.5 ± 0.253.5 ± 2.275.9 ± 0.633.1 ± 2.097.0 ± 0.574.0 ± 0.332.2 ± 2.0
+ FK-ReMask + validity96.5 ± 0.596.5 ± 0.154.0 ± 2.176.3 ± 1.034.2 ± 1.097.4 ± 0.174.5 ± 1.033.5 ± 1.0
+ FK-ReMask + validity + novelty96.6 ± 0.797.4 ± 0.270.8 ± 0.970.1 ± 0.743.0 ± 0.7*97.5 ± 0.468.8 ± 0.942.1 ± 0.4*

 

Changing the novelty reward weight also lets the user set the trade-off between novelty and stability:

 

λ_noveltyvaliduniquenovelstableSUNon-targetstable & on-targetSUN & on-target
076.0 ± 1.996.9 ± 0.852.7 ± 2.276.8 ± 1.933.4 ± 2.196.1 ± 0.974.0 ± 1.932.4 ± 2.1
0.173.8 ± 1.998.6 ± 0.570.9 ± 2.066.6 ± 2.140.6 ± 2.296.7 ± 0.864.5 ± 2.138.9 ± 2.2
0.574.2 ± 1.999.6 ± 0.392.0 ± 1.257.8 ± 2.250.2 ± 2.291.4 ± 1.252.9 ± 2.245.7 ± 2.2
1.072.7 ± 2.099.8 ± 0.296.3 ± 0.851.8 ± 2.248.0 ± 2.286.9 ± 1.546.1 ± 2.243.0 ± 2.2
2.073.4 ± 2.099.6 ± 0.397.1 ± 0.749.8 ± 2.247.1 ± 2.285.2 ± 1.643.6 ± 2.241.2 ± 2.2

 

 

2) reward = evaluator

 

  • Steering with predictor A and evaluate on-target with an independent predictor B.

     

    Following this comment, we evaluated each method’s 512 crystals with three bandgap evaluators. (1) ALIGNN-JARVIS, an ML predictor using the same architecture but different training data; (2) CGCNN, an ML predictor using a different architecture; (3) single-point DFT (Quantum ESPRESSO). FK-ReMask wins over baselines in all cases, with all predictors.

     

    As expected, DFT yields lower on-target rates than ML predictors. Still, the ranking across methods is preserved, and FK-ReMask achieves the highest SUN & on-target yield under every evaluator (Appendix H):

     

On-target comparison table:

methodALIGNN-MP (reward)ALIGNN-JARVISCGCNNSingle-point DFT
MatterGen0.4 ± 0.30.8 ± 0.40.8 ± 0.40.4 ± 0.3
+ Best-of-N7.6 ± 1.24.5 ± 0.94.9 ± 1.01.8 ± 0.6
+ CFG4.9 ± 1.03.7 ± 0.86.8 ± 1.13.3 ± 0.8
+ CFG + Best-of-N37.5 ± 2.113.3 ± 1.513.3 ± 1.59.6 ± 1.3
+ FK-ReMask64.5 ± 2.114.6 ± 1.614.8 ± 1.69.6 ± 1.3

 

SUN & On-target comparison table:

methodALIGNN-MP (reward)ALIGNN-JARVISCGCNNSingle-point DFT
MatterGen0.2 ± 0.20.2 ± 0.20.2 ± 0.20.4 ± 0.3
+ Best-of-N3.9 ± 0.92.0 ± 0.62.5 ± 0.71.0 ± 0.4
+ CFG1.2 ± 0.51.6 ± 0.52.0 ± 0.62.0 ± 0.6
+ CFG + Best-of-N13.9 ± 1.55.9 ± 1.05.3 ± 1.02.9 ± 0.7
+ FK-ReMask32.6 ± 2.17.4 ± 1.27.4 ± 1.23.5 ± 0.8

 

  • DFT single-point+relax a random 30-50 of your "SUN & on-target" hits (and the baselines'): what is the DFT-confirmed rate with cis?

     

    For all SUN & on-target hits among the 512 crystals each method generated (single seed), we computed the DFT bandgap both as a single-point and after a full DFT relaxation, and report the confirmation rate with 95% Wilson intervals (Appendix H, Table A10).

     

    FK-ReMask achieves a relaxed DFT confirmation rate comparable to those of the baselines, with overlapping confidence intervals. Importantly, its larger number of SUN and on-target candidates translates into 14 DFT-confirmed hits, compared with 7 for CFG + Best-of-N, 1 for Best-of-N, and 0 for CFG.

methodSUN & on-target countSingle-point DFT confirm rateRelaxed DFT confirm rate
MatterGen10% [0%, 79%]0% [0%, 79%]
+ Best-of-N205% [1%, 24%]5% [1%, 24%]
+ CFG617% [3%, 56%]0% [0%, 39%]
+ CFG + Best-of-N717% [3%, 15%]10% [5%, 19%]
+ FK-ReMask1677% [4%, 12%]8% [5%, 14%]

 

  • Does τ->0 inflate surrogate-on-target while held-out on-target saturates (ie reward-hacking signature)?

     

    Following the reviewer’s suggestion, for the single-objective bandgap task we evaluated 512 crystals at several selection temperatures using the reward predictor (ALIGNN-MP) and two held-out predictors (ALIGNN-JARVIS and CGCNN). As τ decreases, the reward predictor’s on-target rate does not increase. It is stable and falls slightly at the smallest τ. The held-out rates show no systematic change either. Thus, more aggressive selection does not disproportionately increase the reward predictor’s score.

τALIGNN-MP (reward)ALIGNN-JARVIS (held-out)CGCNN (held-out)
0.263.3 ± 2.117.6 ± 1.717.4 ± 1.7
0.163.9 ± 2.116.8 ± 1.715.8 ± 1.6
0.0561.3 ± 2.218.4 ± 1.713.3 ± 1.5
0.02556.6 ± 2.219.3 ± 1.718.0 ± 1.7

 

  • Head-to-head vs Hasan's resampling-suspension and vs t2c-fk / universal-guidance where they apply.

     

    Following this comment:

     

    • We ran Hasan et al.'s suggestion for collapse remedy: suspending resampling in the final 5% of sampling, in place of re-masking, on the three-objective photovoltaic task. The fraction of SUN crystals meeting all three targets is 1.4 ± 0.5% with suspension, against 2.0 ± 0.6% with selection only and 8.6 ± 1.2% with FK-ReMask (one seed, 512 crystals per method; ± is the binomial standard error). Suspension does not help, because in our case collapse happens at the first selection event: distinct compositions per 8-particle group fall from 7.9 to 2.9 after the first selection event. Re-masking succeeds because it reopens committed atom sites, letting cloned particles diverge again.

       

    • We compared FK-ReMask with Universal Guidance (UG). UG steers only atom positions and the lattice, whereas FK-ReMask also steers atom types, enabling control over properties of the entire crystal. For band gap and bulk modulus tasks, FK-ReMask achieves the highest on-target, stable & on-target, and SUN & on-target yields (single seed, 512 crystals each).

       

      We additionally evaluate a task from the UG paper: generating Li-Co-O crystals with a mean Co-O coordination number of 4. All methods use MatterGen conditioned on Li-Co-O with CFG, which encourages but does not enforce the presence of these elements. Consequently, 80% of vanilla MatterGen's generated crystals contain no cobalt, making the target inaccessible through atomic-position steering alone. UG does not overcome this limitation (only 19% of its crystals contain cobalt), whereas FK-ReMask increases this fraction to 75% and achieves the highest yields on all three metrics.

      ValidUniqueNovelStableSUNOn targetStable & on targetSUN & on target
      Band gap, 2.4 ± 0.1 eV
      MatterGen86.3 ± 1.599.6 ± 0.359.2 ± 2.275.0 ± 1.937.5 ± 2.10.4 ± 0.30.4 ± 0.30.2 ± 0.2
      + Universal Guidance86.1 ± 1.599.8 ± 0.259.6 ± 2.276.8 ± 1.939.6 ± 2.22.7 ± 0.72.1 ± 0.61.2 ± 0.5
      + FK-ReMask80.9 ± 1.796.1 ± 0.968.2 ± 2.182.6 ± 1.751.6 ± 2.264.5 ± 2.151.6 ± 2.232.6 ± 2.1
      Bulk modulus, 150 ± 15 GPa
      MatterGen86.3 ± 1.599.6 ± 0.359.2 ± 2.275.0 ± 1.937.5 ± 2.16.4 ± 1.14.1 ± 0.91.8 ± 0.6
      + Universal Guidance86.5 ± 1.599.4 ± 0.358.4 ± 2.277.7 ± 1.840.0 ± 2.26.1 ± 1.15.5 ± 1.02.0 ± 0.6
      + FK-ReMask76.0 ± 1.996.9 ± 0.852.7 ± 2.276.8 ± 1.933.4 ± 2.196.1 ± 0.974.0 ± 1.932.4 ± 2.1
      Mean Co–O coordination, Li–Co–O, 4 ± 0.5
      MatterGen, conditioned on Li–Co–O48.8 ± 2.260.9 ± 2.278.9 ± 1.837.7 ± 2.19.8 ± 1.35.1 ± 1.02.9 ± 0.82.0 ± 0.6
      + Universal Guidance42.0 ± 2.265.4 ± 2.183.4 ± 1.629.1 ± 2.010.9 ± 1.44.5 ± 0.90.8 ± 0.40.8 ± 0.4
      + FK-ReMask74.4 ± 1.960.7 ± 2.281.4 ± 1.738.9 ± 2.213.3 ± 1.537.1 ± 2.19.8 ± 1.35.7 ± 1.0

       

    • We compared FK-ReMask with T2C-FK, the particle sampler of ALM. T2C-FK also resamples particles based on a reward on the clean estimate, but it resamples only when the effective sample size drops, and it has no re-masking, so atom types committed early cannot be revised. We do not use ALM's language model as a conditioner, because it was not trained to generate crystals at a numeric property value such as a band gap of 2.4 eV. FK-ReMask achieves a higher SUN & on-target rate (over 3 seeds):

      methodvaliduniquenovelstableSUNon-targetstable & on-targetSUN & on-target
      MatterGen86.0 ± 0.899.7 ± 0.159.7 ± 0.673.3 ± 1.437.0 ± 1.71.2 ± 0.40.8 ± 0.30.5 ± 0.2
      + Best-of-N82.4 ± 0.3100.0 ± 0.070.4 ± 1.469.2 ± 2.041.9 ± 0.98.7 ± 0.66.1 ± 0.44.2 ± 0.2
      + T2C-FK81.1 ± 0.999.2 ± 0.369.1 ± 1.073.0 ± 1.043.7 ± 0.526.2 ± 1.319.2 ± 1.013.7 ± 0.3
      + FK-ReMask81.2 ± 0.396.5 ± 0.570.1 ± 1.283.8 ± 0.654.8 ± 1.763.9 ± 0.553.6 ± 1.135.4 ± 1.7
Reviewer_GX6XOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper introduces FK-ReMask, an inference-time, gradient-free approach for steering pretrained crystal diffusion models toward desired material properties. The method combines Feynman–Kac particle selection with atom re-masking, enabling optimization of both continuous structural variables and discrete chemical compositions without retraining. The authors demonstrate improvements over baseline methods across four individual material properties and jointly optimize three photovoltaic properties. The results show improved yields of stable, unique, novel, and property-compliant materials.

Critical strengths

1. Multi-property optimization is a compelling problem. Simultaneously targeting multiple material properties, while maintaining stability and novelty, is an important challenge in materials discovery. I like the ability to modify target specifications without retraining the generative model.

2. Potential value in data-limited settings. The proposed approach is particularly attractive when property-labeled training data are scarce or expensive to obtain. The combination of particle selection and re-masking is interesting, and the demonstrated improvements over baseline methods are encouraging.

Critical weaknesses

1. Unclear value proposition when large datasets are available. The primary advantage appears to be avoiding retraining. However, when sufficiently large property-labeled datasets are available, it is unclear whether inference-time steering offers meaningful advantages over directly conditioned or fine-tuned generative models. I would encourage the authors to clarify the regimes in which their approach is preferable, particularly as a function of training-data availability.

2. Sensitivity to the cost of property evaluation. The method repeatedly evaluates property predictors during generation. How does its computational efficiency scale with the cost of these forward evaluators? This becomes particularly important when the desired properties require expensive simulations or high-fidelity calculations rather than inexpensive ML surrogates. A cost-versus-performance analysis would help establish the practical utility of the approach.

Questions for authors

See weakness section

2 comments

Thank you for highlighting the importance of the problem, that the solution is interesting, and that the empirical improvements are encouraging.

1) Value of our method when large labeled datasets are available.

When abundant labeled data are available for a single, fixed target, a conditional or fine-tuned model is indeed a strong option. We added this point to the intro and discuss it in the Limitations. Our method offers several advantages even when large training datasets are available:

  • Multiple properties with limited joint labels. Fine-tuning a conditional model on several properties typically needs crystals that are annotated with all labels, but label combinations often do not cover all desired combinations:

     

    In the JARVIS-DFT dataset, each criterion of our three-objective task (bandgap within 1.3 ± 0.1 eV, SLME ≥ 25%, hole effective mass between 0 and 0.3) is met by 781 to 2,203 crystals, each pair by 2 to 202, and all three by none of the 2,332 crystals that carry all three labels. Alex-MP-20, MatterGen's training set, has no SLME or hole effective mass labels at all.

     

    Our method combines independently trained property predictors. Thus, it does not require a dataset of crystals jointly labeled with all target properties.

# objectivesbandgapSLMEhole massbulk modulusAlex-MP-20: #on-target (#total labelled)JARVIS: #on-target (#total labelled)
1✓904 (43,295)1,064 (75,993)
1✓02,203 (9,770)
1✓0781 (3,568)
1✓341 (5,526)2,354 (23,824)
2✓✓0202 (9,770)
2✓✓02 (3,568)
2✓✓21 (5,526)69 (23,824)
2✓✓0159 (2,332)
2✓✓0182 (6,595)
2✓✓032 (2,140)
3✓✓✓00 (2,332)
3✓✓✓015 (6,595)
3✓✓✓01 (2,140)
3✓✓✓018 (1,817)
4✓✓✓✓00 (1,817)

 

  • Controlling trade-offs between objectives at inference time. Targeting a property can degrade existing capabilities of the generative model, such as validity and novelty, and several target properties can conflict with one another.

     

    FK-ReMask lets users control these trade-offs at inference time by adding reward terms and adjusting their relative weights. Fine-tuning, in contrast, requires updating the model weights for each new or changed objective, and potentially compromising previously learned capabilities (catastrophic forgetting).

     

    For example, steering MatterGen with FK-ReMask toward a bulk modulus target lowered validity from 86.0% to 76.6%. Adding a validity term to the reward raised it to 96.5% while the on-target rate was preserved (97.4%) (Appendix E, Table A4).

     

    On the same property target, the weight of a novelty term sets the balance between novelty and stability: at weights 0, 0.1, 0.5 and 1.0, novelty increased (52.7% → 70.9% → 92.0% → 96.3%) while stability decreased (76.8% → 66.6% → 57.8% → 51.8%) (Appendix E, Table A5).

     

  • Complementary to conditional models. We applied FK-ReMask on top of MatterGen with classifier-free guidance (CFG), whose property adapter is fine-tuned on labelled data. On bandgap and bulk modulus, it improves the SUN & on-target rates in comparison to CFG and CFG + Best-of-N (Appendix G).

propertymethodvaliduniquenovelstableSUNon-targetstable & on-targetSUN & on-target
Bandgap (E_g)CFG86.3 ± 0.299.0 ± 0.269.0 ± 1.666.9 ± 1.237.7 ± 0.65.0 ± 0.33.3 ± 0.11.5 ± 0.3
CFG + Best-of-N86.0 ± 1.197.7 ± 0.268.9 ± 1.568.9 ± 1.039.5 ± 0.436.5 ± 0.925.2 ± 0.913.5 ± 0.4
CFG + FK-ReMask88.2 ± 1.590.7 ± 0.262.1 ± 1.086.1 ± 0.9*47.3 ± 0.5*63.3 ± 0.4*55.7 ± 0.8*29.9 ± 0.3*
Bulk Modulus (K)CFG81.2 ± 1.499.5 ± 0.157.1 ± 0.866.1 ± 1.327.9 ± 1.153.3 ± 0.336.8 ± 0.515.8 ± 0.2
CFG + Best-of-N78.7 ± 0.698.8 ± 0.255.2 ± 0.165.9 ± 1.327.3 ± 0.799.7 ± 0.165.7 ± 1.427.2 ± 0.8
CFG + FK-ReMask80.8 ± 0.595.3 ± 0.552.7 ± 1.576.4 ± 0.7*33.7 ± 1.499.7 ± 0.276.2 ± 0.6*33.5 ± 1.5*

 

2) Sensitivity to the cost of property evaluation.

  • Using computationally expensive evaluators such as DFT during generation is indeed impractical. However, we view FK-ReMask as the first stage of a multi-stage materials discovery pipeline, where fast ML predictors guide the search toward promising candidates for further expensive DFT calculations and experimental validation.

     

    This follows the principles discussed by Miracle et al. (2017) [1]. They show that relying only on expensive evaluation methods leaves most of the search space unexplored, noting that after 12 years, only 122 new alloy materials have been made out of approximately 171 million possible candidates, representing just 7.2 × 10⁻⁷ of the search space. This highlights the importance of fast, even if imperfect, screening to explore a much larger part of the chemical space.

     

    To assess whether our fast screening method (FK-ReMask) translates to more expensive evaluation, we further evaluated all 512 generated crystals per method with single-point DFT in Quantum ESPRESSO. FK-ReMask has the highest SUN & on-target yield under DFT: 3.5%, against 2.9% for CFG + Best-of-N, 2.0% for CFG, 1.0% for Best-of-N, and 0.4% for MatterGen (Appendix H, Table A8).

     

  • A cost-versus-performance analysis.. We consider the cost to be the number of ML-based bandgap predictor calls. In a single crystal generation with P=1 target properties, K=8 trajectories, and E=35 selection events, our method calls the property predictor P×K×(E+1) = 1 × 8 × 36 = 288 times (the E selection events plus the final pick). We vary E to get different numbers of predictor calls and calculate the number of on-target and SUN & on-target crystals.

     

    In this case, halving the predictor calls, from 288 to 144, reduces the on-target rate (64.5% to 56.6%) but leaves the SUN & on-target yield nearly unchanged (32.6% to 31.6%). A further reduction to 64 calls, about 4.5-fold fewer, lowers SUN & on-target to 26.8%, still about twice that of CFG + Best-of-N (13.9%).

     

    While DFT-level evaluation remains impractical during generation, this analysis supports extending FK-ReMask to more expensive evaluators, such as larger ML models or ensembles.

MethodPredictor callsOn-target (%)SUN & on-target (%)
FK-ReMask, E = 7056868.236.1
FK-ReMask, E = 35 (default)28864.532.6
FK-ReMask, E = 1714456.631.6
FK-ReMask, E = 76443.226.8
FK-ReMask, E = 33234.621.7

 

[1] D. B. Miracle, B. Majumdar, K. Wertz, and S. Gorsse, "New strategies and tests to accelerate discovery and development of multi-principal element structural alloys," Scripta Materialia, vol. 127, pp. 195–200, 2017.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.