acceptodds
Under review as a conference paper at ICLR 2027

Transferable Boltzmann Generator for Direct Air Capture

Abstract

Screening metal–organic frameworks (MOFs) for direct air capture requires efficient estimation of dilute-limit CO adsorption. Conventional Widom insertion spends many energy evaluations on unfavorable configurations, limiting the scale of screening with machine-learned interatomic potentials (MLIPs). We introduce a transferable MOF-conditioned Boltzmann generator for estimating the Henry coefficient without per-material retraining. The model uses equivariant MOF features from a pretrained MLIP to construct a normalized proposal over CO position, orientation, and internal deformation, enabling importance-weighted estimation with flexible molecules. Periodically wrapped Gaussian position mixtures and conditional orientation and deformation mixtures support direct sampling and density evaluation, while MOF features are computed once and reused across draws. Across unseen MOFs spanning weak to strong adsorption, our model maintains high sampling efficiency and closely matches reference configuration and energy distributions after reweighting. It estimates flexible-gas Henry coefficients at roughly 700-fold lower inference cost than rigid-gas Widom insertion while capturing systematic effects of molecular deformation that additional rigid-gas sampling cannot recover.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
boltzmann generator, metal-organic framework, free energy estimate, material discovery

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_eXaZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces a MOF-conditioned Boltzmann generator for dilute-limit CO₂ adsorption calculations. The model parameterizes a normalized proposal over molecular position, orientation and internal deformation, using pretrained MLIP features and a shared position–orientation score. Importance weighting enables Henry-coefficient estimation under the chosen MLIP. Experiments demonstrate substantial sampling gains across held-out MOFs.

Critical strengths
  1. The problem is well motivated. Uniform insertion wastes expensive MLIP evaluations on unfavorable configurations. Learning a transferable proposal directly addresses this bottleneck.

  2. The probabilistic formulation is technically careful. The method explicitly handles periodicity, coordinate Jacobians and oxygen-exchange symmetry, which are essential for valid density evaluation and importance weighting.

  3. The computational results are promising. A shared model transfers to 672 held-out MOFs without per-material retraining and achieves high sampling efficiency across adsorption-strength quartiles. Caching material features makes repeated sampling inexpensive.

  4. The paper appropriately acknowledges that efficient guest sampling does not establish converged adsorption estimates for fully flexible frameworks.

Critical weaknesses
  1. Methodological novelty is limited and insufficiently isolated. Conditional density factorization, mixture distributions, pretrained equivariant features and importance sampling are established tools. The main task-specific contribution appears to be their integration, particularly the shared position–orientation score. However, the ablations do not directly test this coupling or compare against a simpler normalized joint pose model. Consequently, the necessity of the distinctive design remains unclear.

  2. The transfer evaluation is relatively permissive. Splitting by MOF identifier does not establish generalization across structural families, linkers or topologies. All test metals occur in training, and dataset filtering restricts the range of framework sizes and pore accessibility.

  3. Reference reliability needs stronger independent validation. Flexible-gas accuracy is evaluated on 594 of 672 MOFs passing a reference-precision threshold. Excluded structures may represent particularly difficult cases. Incorporating model-discovered sites into the reference improves coverage, but particle-bootstrap intervals cannot capture uncertainty from undiscovered modes. Independent sampling checks would strengthen the evidence.

  4. The headline speedup does not establish total workflow savings. Approximately 700× acceleration corresponds to the 16-draw operating point. Training-data generation and model training are excluded, and the more accurate configurations have different cost ratios. Equal-error comparisons and amortization analysis are needed for realistic screening workloads.

Questions for authors
  1. How much does the shared position–orientation score improve performance over independently predicted factors or a simpler conditional mixture?

  2. How does performance change under linker-, topology- or structural-family-based splits?

  3. What is the results on reference-unqualified MOFs, and can independent sampling validate dominant-site coverage?

  4. At matched estimation error, how many new MOFs must be screened to recover preparation and training costs?

Can the authors address the novelty and generalization of the method used in this paper?

3 comments
AuthorsOpenReview ↗

Response to Reviewer eXaZ (1/2)

We appreciate the reviewer’s suggestions for isolating the architectural contribution and strengthening the generalization, reference-validation, and cost analyses. We provide new shared-score ablations, OOD evaluations, improved model-free AIS references, and a workflow cost analysis. Manuscript Figure 1 and Table 2 reflect the updated evaluation. The AIS reference-generation protocol is provided in Appendix F. Manuscript changes are highlighted in blue. Additional experimental results will be incorporated into the revised manuscript.

W1 & Q1. What is the methodological novelty, and how much does the shared pose score improve over simpler normalized pose models?

We tested the contribution of the shared position–orientation score against two simpler normalized alternatives, retraining each with the same protocol.

  • (A) Independent factors. Position weights use only the per-voxel logit, πj∝ebj\pi_j\propto e^{b_j}. The orientation distribution retains the pose score SS, but position weights no longer account for which orientations are favorable.
  • (B) Simpler conditional mixture. We remove the pose score entirely. Position weights remain πj∝ebj\pi_j\propto e^{b_j}, while orientation weights over the same 256 axes are predicted by an MLP from host features at the sampled position.

Results on the test MOFs are shown below.

ModelMean nESS (%), Q1 / Q2 / Q3 / Q4 ↑Mean nESS (%), all ↑Mean ∣Δln⁡KH∣\lvert\Delta\ln K_H\rvert ↓
Flexible CO₂ (668 MOFs)
Full model (shared score)40.3 / 63.5 / 59.1 / 44.952.00.031
(A) Independent factors21.1 / 39.7 / 32.3 / 22.729.00.037
(B) Simple conditional mixture5.7 / 9.0 / 5.9 / 3.36.00.076
Rigid CO₂ (669 MOFs)
Full model (shared score)43.3 / 65.7 / 64.7 / 52.456.50.021
(A) Independent factors22.1 / 41.7 / 35.4 / 26.531.40.026
(B) Simple conditional mixture6.0 / 9.8 / 6.5 / 3.86.50.056

All evaluations use 32,768 samples per MOF, with Q1–Q4 denoting adsorption-strength quartiles. The shared score improves sampling efficiency in every quartile for both flexible and rigid CO₂. Overall, it achieves approximately 1.8×1.8 \times the mean nESS of (A) and 8.7×8.7 \times that of (B). These ablations isolate the contribution of the shared score beyond the use of normalized mixture distributions alone.

W2 & Q2. Does the method generalize across MOF families, linkers, and topologies under more demanding train–test splits?

We added extrapolation tests by excluding selected metals, linkers, or topologies from training and retraining with the same protocol. We selected frequent groups in our dataset and evaluated on the corresponding original test MOFs with reference bootstrap uncertainty ≤0.2\leq 0.2.

CategoryGroup excluded from trainingTest MOFsMean ∣Δln⁡KH∣\lvert\Delta\ln K_H\rvert ↓Mean nESS (%) ↑
MetalCu1030.027 → 0.02659.5 → 51.3
MetalCd590.023 → 0.02455.7 → 50.4
MetalCo800.022 → 0.02848.2 → 40.9
Linker4,4′-bipyridine330.025 → 0.02568.3 → 66.8
LinkerTerephthalate (BDC)200.027 → 0.02763.0 → 62.0
Topologysql760.024 → 0.02557.2 → 54.2
Topologyhcb300.044 → 0.04849.6 → 48.6
Topologypcu170.018 → 0.01764.0 → 61.0

Arrows compare the model trained on all groups with the model retrained after excluding the specified group. Accuracy and sampling efficiency are largely maintained for unseen linkers and topologies. Metal exclusion additionally requires generalization to elements absent from the proposal’s training data and causes a larger reduction in sampling efficiency.

For size extrapolation, we randomly selected 200 additional MOFs with 201–300 atoms per unit cell using the same pore-accessibility criteria and generated new references. The model trained on MOFs with at most 200 atoms was evaluated without retraining.

Test setAtoms per unit cellTest MOFsMean ∣Δln⁡KH∣\lvert\Delta\ln K_H\rvert ↓Mean nESS (%) ↑
Original test set≤2006680.03152.0
Size OOD201–3002000.03561.7

All evaluations use 32,768 samples per MOF and references with bootstrap uncertainty ≤0.2\leq 0.2. The model maintains strong accuracy and sampling efficiency on larger MOFs.

Together, these results demonstrate that the learned proposal generalizes beyond the original training distribution, maintaining strong accuracy and sampling efficiency across unseen metal elements, linkers, topologies, and larger frameworks without adapting the model to the test MOFs.

AuthorsOpenReview ↗

Response to Reviewer eXaZ (2/2)

W3 & Q3. How reliable are the reference estimates, how does the method perform on reference-unqualified MOFs, and can independent sampling validate site coverage?

We agree that particle-bootstrap precision alone cannot establish reference convergence, because it cannot detect entirely missed adsorption sites. After submission, we regenerated the references using a model-free initialization built from Widom insertions and geometry-optimized adsorption sites, removing dependence on our learned proposal.

For each MOF, we selected 1,024 distinct low-energy configurations from 1M Widom insertions and optimized them to locate adsorption minima. We supplemented these with optimized configurations from a preliminary model-free AIS run and constructed Hessian-based local Gaussian proposals around the discovered sites. Appendix F, “AIS reference generation,” details the initialization, annealing schedule, production sampling, and bootstrap intervals.

After submission, we also generated new sequential Monte Carlo (SMC) training data with 4,096 flexible CO₂ configurations per training MOF and retrained the same architecture. Figure 1 and Table 2 report the retrained model against the new references.

Metric across 672 MOFs, flexible CO₂Original referenceUpdated reference
MOFs with bootstrap uncertainty ≤0.1\leq 0.1 ↑405647
MOFs with bootstrap uncertainty ≤0.2\leq 0.2 ↑594668
MOFs with Emin⁡ref−Emin⁡opt≤kBTE_{\min}^{\mathrm{ref}}-E_{\min}^{\mathrm{opt}}\leq k_BT ↑637672

Here, Emin⁡refE_{\min}^{\mathrm{ref}} is the lowest sampled energy and Emin⁡optE_{\min}^{\mathrm{opt}} is the lowest energy found by geometry optimization. These energy comparisons assess whether sampling reaches the discovered low-energy regions, independently of bootstrap precision.

For reference-unqualified MOFs, accuracy cannot be reliably assessed because the reference values themselves remain insufficiently resolved. We therefore provide updated evaluations using the more reliable references, expanding the KHRFK_H^{RF} evaluation set from 594 to 668 MOFs.

W4 & Q4. At matched estimation error, how many new MOFs must be screened to recover data-generation and training costs?

We agree that inference speedup alone does not establish total workflow savings. We therefore estimated the one-time preparation cost and the number of new MOFs needed to recover it at comparable estimation error.

Preparation stage, 6,052 training MOFsGPU-hours
Widom screening, 100,000 insertions per MOF649
Geometry optimization of adsorption sites407
Site Hessian calculations≈32
SMC training-data generation180
Feature extraction≈1
Model training17
Total≈1,290

Preparation costs from a single H200 GPU.

Approximate mean ∣Δln⁡KH∣\lvert\Delta\ln K_H\rvertOurs, draws / GPU-s per MOFWidom-initialized AIS, particles / GPU-s per MOFInference speedupBreak-even new MOFs
0.1132 / 0.504,096 / 1,075≈2,100×≈4,300
0.0864 / 0.618,192 / 1,750≈2,900×≈2,700

The comparison uses flexible CO2 on 665 reference-qualified test MOFs against the updated model-free references. Inference costs use the H200 timing basis of Figure 1. Ours includes feature and context construction, and the AIS baseline includes its initial 100,000 Widom insertions.

We calculate the break-even point as Nbreak−even=⌈Cprep/(cAIS−cours)⌉N_{\mathrm{break-even}}=\lceil C_{\mathrm{prep}}/(c_{\mathrm{AIS}}-c_{\mathrm{ours}})\rceil, where CprepC_{\mathrm{prep}} is the one-time preparation cost and cc is the inference cost per new MOF. Screening approximately 2,700–4,300 additional MOFs therefore recovers preparation and training costs at these accuracy levels. Larger campaigns benefit from applying the shared model without per-MOF retraining.

The extrapolation results in Q2 further support reuse of the trained model across new metals, linkers, topologies, and larger frameworks. The preparation cost can therefore be amortized across successive screening campaigns rather than incurred separately for each new MOF or test set, provided the target energy function and thermodynamic conditions remain unchanged.

Dear Reviewer eXaZ,

Thank you again for your thoughtful review. We recently posted our response and additional experimental results. With the author-reviewer discussion period closing today (~ 13hr left), we would appreciate knowing whether these have addressed your concerns, or whether any points still need clarification.

Thank you

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

This paper introduces a transferable, MOF-conditioned Boltzmann generator for estimating dilute-limit CO₂ adsorption under a fixed machine-learned interatomic potential. The model constructs a normalized proposal over molecular position, orientation, and internal deformation using periodically wrapped Gaussian mixtures, Watson mixtures, and conditional Gaussian mixtures. A shared pose score couples position and orientation, while cached equivariant features from a pretrained MLIP reduce the cost of generating additional samples. Importance weighting then estimates the configurational integrals required for Henry coefficients.

The authors conduct experiments on 6,724 MOFs spanning a wide range of adsorption strengths and compare with conventional Widom insertion with optional annealed importance sampling. The experiments demonstrate substantial improvements over baselines.

Critical strengths
  1. The method provides a normalized density over complete molecular configurations, enabling importance weighting rather than relying solely on predicted adsorption densities. The treatment of periodic positions, molecular coordinate transformations, and oxygen-exchange symmetry is detailed and technically meaningful.
  2. The evaluation covers hundreds of held-out MOFs and separates results by adsorption-strength quartile. Table 2 provides convincing evidence of practical improvement at the reported operating points.
Critical weaknesses

Transferability is demonstrated within a filtered structural distribution, with limited evidence of generalization across distinct MOF families.The dataset is restricted by pore accessibility, unit-cell size, and successful identification of at least one negative-interaction-energy configuration in a large Widom sweep. These choices define a useful benchmark, but narrow the population to which the results apply. A grouped split by framework or linker family, or an analysis of performance versus nearest-training-structure similarity, would clarify the strength of transfer. Performance on the structures excluded by reference uncertainty should also be characterized separately.

Questions for authors
  1. How is the AIS reference initialized and correctly weighted? 2.What existing evidence supports reference convergence beyond particle-bootstrap precision?
3 comments
AuthorsOpenReview ↗

Response to Reviewer wShY (1/2)

We appreciate the reviewer’s comments on transferability and reference reliability. We have added OOD evaluations across metals, linkers, topologies, and framework sizes, and recomputed the AIS references with improved adsorption-site coverage. Manuscript Figure 1 and Table 2 reflect the updated evaluation. The AIS reference-generation protocol is provided in Appendix F. Manuscript changes are highlighted in blue. Additional experimental results will be incorporated into the revised manuscript.

W1. Does transferability extend beyond the filtered MOF distribution, and how does the method perform on reference-excluded structures?

We added tests of generalization to unseen metals, linkers, topologies, and larger unit cells. For the first three, we selected frequent groups in our dataset, removed all training MOFs containing each group, and retrained with the same protocol. We evaluated on the corresponding original test MOFs with reference bootstrap uncertainty ≤0.2\leq 0.2.

CategoryGroup excluded from trainingTest MOFsMean ∣Δln⁡KH∣\lvert\Delta\ln K_H\rvert ↓Mean nESS (%) ↑
MetalCu1030.027 → 0.02659.5 → 51.3
MetalCd590.023 → 0.02455.7 → 50.4
MetalCo800.022 → 0.02848.2 → 40.9
Linker4,4′-bipyridine330.025 → 0.02568.3 → 66.8
LinkerTerephthalate (BDC)200.027 → 0.02763.0 → 62.0
Topologysql760.024 → 0.02557.2 → 54.2
Topologyhcb300.044 → 0.04849.6 → 48.6
Topologypcu170.018 → 0.01764.0 → 61.0

Arrows compare the model trained on all groups with the model retrained after excluding the specified group. Accuracy and sampling efficiency are largely maintained for unseen linkers and topologies. Excluding metals causes a larger efficiency reduction because it requires extrapolation to elements absent from the proposal’s training data.

For size extrapolation, we randomly selected 200 additional MOFs with 201–300 atoms per unit cell using the same pore-accessibility criteria and generated new references. The model trained on MOFs with at most 200 atoms was evaluated without retraining. The model maintains strong accuracy on larger MOFs.

Test setAtoms per unit cellTest MOFsMean ∣Δln⁡KH∣\lvert\Delta\ln K_H\rvert ↓Mean nESS (%) ↑
Original test set≤2006680.03152.0
Size OOD201–3002000.03561.7

The pore-accessibility criteria are motivated by the physical requirements of direct air capture. CO₂ must be able to enter the framework pores to access internal adsorption sites, so these filters select materials relevant to the application rather than being tuned to favor our model.

For MOFs whose references have not converged sufficiently, unreliable target values make accuracy evaluation impossible. We therefore regenerated the references after submission, improving reference reliability and increasing the number of MOFs meeting the bootstrap uncertainty threshold from 594 to 668. We describe the updated reference generation and additional convergence checks in Q1–Q2.

AuthorsOpenReview ↗

Response to Reviewer wShY (2/2)

Q1 & Q2. AIS reference generation and convergence

In our original submission, we initialized AIS from Widom- (qWq_W) and model-discovered sites (qMq_M) to cover adsorption sites missed by Widom-initialized AIS (Figure 3), with a uniform component (qUq_U) preserving pose support.

=0.9\left[\tfrac12q_{\mathrm W}^{100\mathrm{k}}(X)+\tfrac12q_{\mathrm M}(X)\right] +0.1q_{\mathrm U}(X),$$ where $X$ denotes CO₂ coordinates. The proposals $q_{\mathrm W}$ and $q_{\mathrm M}$ use wrapped position Gaussians and symmetric von Mises–Fisher orientation kernels centered on Widom- and model-discovered poses, respectively. The proposal $q_{\mathrm U}$ uses uniform positions and orientations. Samples drawn from $q_0$ have unit initial weights. Along $f_\lambda=q_0^{1-\lambda}e^{-\lambda\beta\Delta U}$, where $\Delta U$ includes the gas internal energy and gas–MOF interaction energy, the weight increment is $\Delta\log w=(\lambda_k-\lambda_{k-1})[-\beta\Delta U(X)-\log q_0(X)]$, evaluated using the full mixture density. For each MOF, a pilot selects the $\lambda$ schedule at a conditional effective sample size fraction of 0.75 and tunes move scales. Fresh production particles follow the frozen schedule and scales, excluding pilot particles. At each intermediate stage, weights are updated before five Metropolis–Hastings moves targeting $f_\lambda$. Particle bootstrapping quantifies sampling uncertainty but cannot detect entirely missed adsorption sites. After submission, we therefore regenerated the references for all 672 MOFs using a model-free initialization informed by geometry optimization. For each MOF, we performed 1M additional Widom insertions and selected 1,024 starting configurations in order of increasing interaction energy, removing similar positions and orientations. We geometry-optimized these configurations and expanded the site set by optimizing high-weight configurations from a preliminary model-free AIS run initialized around Widom poses and the optimized minima. After removing duplicate minima, we computed site Hessians to construct local Gaussian proposals with covariance matrices derived from their curvatures. Their mixture $q_{\mathrm L}$ enters the updated initialization as $$q_0^{\mathrm{new}}(X) =\tfrac12\left[0.9q_{\mathrm W}^{1\mathrm{M}}(X)+0.1q_{\mathrm U}(X)\right] +\tfrac12q_{\mathrm L}(X).$$ Appendix F, “AIS reference generation,” details the initialization, annealing schedule, production sampling, and bootstrap intervals. After submission, we also generated new sequential Monte Carlo (SMC) training data with 4,096 flexible CO₂ configurations per training MOF and retrained the same architecture. Figure 1 and Table 2 report the retrained model against the new references. | Metric across 672 MOFs | Original $K_H^{\mathrm{RR}}$ | Updated $K_H^{\mathrm{RR}}$ | Original $K_H^{\mathrm{RF}}$ | Updated $K_H^{\mathrm{RF}}$ | | --- | ---: | ---: | ---: | ---: | | MOFs with bootstrap uncertainty $\leq 0.1$ ↑ | 574 | 642 | 405 | 647 | | MOFs with bootstrap uncertainty $\leq 0.2$ ↑ | 661 | 669 | 594 | 668 | Beyond bootstrap precision, we compared sampled energies with the lowest energies found by geometry optimization ($E_{\min}^{\mathrm{opt}}$). For $K_H^{RF}$, the number of MOFs satisfying $E_{\min}^{\mathrm{ref}}-E_{\min}^{\mathrm{opt}}\leq k_BT$ increased from 637 to 672 after regenerating new AIS reference.These checks support improved coverage and stability.

Dear Reviewer wShY,

Thank you again for your thoughtful review. We recently posted our response and additional experimental results. With the author-reviewer discussion period closing today (~ 13hr left), we would appreciate knowing whether these have addressed your concerns, or whether any points still need clarification.

Thank you

Reviewer_DpT9OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Overall, the paper is a well executed application paper. It takes a complex technical problem with important significance for the future of the planet and uses some useful ML methods to improve performance over other recent approaches. They learn a sophisticated proposal distribution, by leveraging a sophisticated pretrained encoder, that can be used for estimating system properties of their target physical systems with better efficiency than existing computational approaches.

Critical strengths

The paper provides an elegant solution to a tricky problem. They leverage an existing domain-specific encoder for these molecular systems to train a little generator that can be used in an importance sampler to estimate a high-value, but expensive to compute, quantity. The exposition on the problem background is good and the solution is well motivated with good results.

Critical weaknesses

My primary critique is a bit pedantic. I feel that 'Boltzmann Generator' is a misleading name for the work. The paper estimates an expectation (eq 1) using importance sampling and sequential Monte Carlo. Upon seeing the title, I expected that the proposal distribution for these would have been given by a learned Boltzmann distribution, as has been done in other ML applications. Learning such distributions requires tricky learning approaches such as score matching. Instead, however, the paper's proposal distribution (sec 3.2) is a closed-form conditional distribution with tractable density. The parameters of this distribution come from a sophisticated neural network applied to the conditioning inputs, but it can be learned using straightforward MLE training on a set of labeled examples from the target conditional distribution. Can you please clarify the terminology? I understand that Boltzmann distributions appear elsewhere in the paper, regarding the physical systems under study, but the ML component seems to not involve a Boltzmann distribution.

Questions for authors

Is there a natural extension of your work to an approach that doesn't learn a generator that can be used to sample independently from a proposal distribution, but instead can be used inside an MCMC scheme?

You use an expensive computational simulation to provide ground truth target labels. Are both the expensive approach and your proposed fast approach unbiased estimators? Otherwise, I worry that using the simulation data may be unreliable.

You split existing MOF data into training and evaluation sets. Can you describe how you split these sets in order to ensure that there is an appropriate level of extrapolation required to generalize from training to evaluation? I can imagine a uniform splitting approach could be unwise.

I get the impression that a key result of your paper is that the sampler can generalize across MOFs. Are there other papers working on related problems (e.g. with learned force fields) that have also thought about such extrapolation? I don't have any sense of whether it is difficult or standard. How impressive is your extrapolation result?

4 comments
AuthorsOpenReview ↗

Response to Reviewer DpT9 (1/3)

We appreciate the reviewer’s thoughtful feedback and recognition of the practical value of our approach. We have recomputed the AIS references with improved adsorption-site coverage and updated manuscript Figure 1 and Table 2 accordingly. We also provide new OOD evaluations and results using the learned proposal within MCMC. The AIS reference-generation protocol is provided in Appendix F. Manuscript changes are highlighted in blue. Additional experimental results will be incorporated into the revised manuscript.

W1. Is the term “Boltzmann generator” appropriate for a model trained by MLE?

We appreciate the opportunity to clarify our terminology. In our understanding, MLE training on examples from the target conditional distribution is compatible with the term “Boltzmann generator,” which describes the distribution being targeted rather than a particular training objective. While neural samplers [1, 2] use energy-function evaluations for training, the original BG paper [3] includes maximum-likelihood training by example, and PROSE [4] trains a BG by maximum likelihood on molecular simulation data. Our method follows this usage. We learn a proposal by MLE and apply importance weighting against the physical Boltzmann target.

[1] Zhang, Q. and Chen, Y. Path Integral Sampler: A Stochastic Control Approach for Sampling. ICLR, 2022.

[2] Havens, A., Karrer, B., and Shaul, N. Flow Sampling: Learning to Sample from Unnormalized Densities via Denoising Conditional Processes. ICML, 2026.

[3] Noé, F., Olsson, S., Köhler, J., and Wu, H. Boltzmann Generators: Sampling Equilibrium States of Many-Body Systems with Deep Learning. Science, 2019.

[4] Tan, C. B. et al. Amortized Sampling with Transferable Normalizing Flows. NeurIPS, 2025.

Q1. Can the learned generator be used within MCMC instead of only generating independent proposal samples?

Yes. Since our model provides both samples and their normalized density, it can be used as an independence Metropolis–Hastings proposal. We explored this approach early in the project but omitted detailed results because of its higher GPU cost. We now report an evaluation across all 672 test MOFs using AIS initialized from the learned proposal, combining local MCMC moves with independence proposals from the model.

MethodSamples / particlesTarget CESSMCMC moves per stageMean ∣Δln⁡KH∣\lvert\Delta\ln K_H\rvert ↓Mean GPU time (s/MOF) ↓
Direct IS4,096——0.04204.82
Learned-proposal AIS2560.7550.07344.87
Learned-proposal AIS1,0240.7550.052916.56
Learned-proposal AIS1,0240.95100.041388.38
Learned-proposal AIS4,0960.95100.0371348.09

CESS denotes conditional effective sample size, and its target fraction controls the spacing between successive AIS distributions. Errors are measured against the updated references described in Q2 (see below).

The learned proposal thus supports MCMC-based sampling, but the additional refinement yields only a modest reduction in mean error at substantially greater GPU cost. Since our goal is a practical replacement for conventional sampling methods, we prioritize accuracy per GPU hour over accuracy per energy evaluation.

AuthorsOpenReview ↗

Response to Reviewer DpT9 (2/3)

Q2. Are both the simulation-based and proposed estimators unbiased, and can the simulation data be trusted?

Yes, both our model and AIS provide unbiased estimates. Training on simulation data does not bias our importance-weighted estimates. We use these data to learn a proposal with full configurational support, and importance weights correct for its discrepancy from the target Boltzmann distribution. With exact gas normalization, the resulting estimate is unbiased.

The key issue is convergence rather than unbiasedness. Neither unbiasedness nor training directly against the energy function guarantees convergence at a finite sampling budget. Our learned proposal therefore aims to place probability mass in thermodynamically important regions, reducing importance-weight variance.

Particle bootstrapping quantifies sampling uncertainty but cannot detect entirely missed adsorption sites. After submission, we therefore performed additional geometry optimization to identify local minima and incorporated the discovered sites into an initialization for the updated AIS references. For each MOF, we selected 1,024 distinct low-energy configurations from 1M Widom insertions as starting points for geometry optimization. We then constructed local Gaussian proposals around the resulting adsorption sites using their Hessians and generated new AIS references. Appendix F, “AIS reference generation,” details the initialization, annealing schedule, production sampling, and bootstrap intervals.

After submission, we also generated new sequential Monte Carlo (SMC) training data with 4,096 flexible CO₂ configurations per training MOF and retrained the same architecture. Figure 1 and Table 2 report the retrained model against the new references.

Metric across 672 MOFsOriginal KHRRK_H^{\mathrm{RR}}Updated KHRRK_H^{\mathrm{RR}}Original KHRFK_H^{\mathrm{RF}}Updated KHRFK_H^{\mathrm{RF}}
MOFs with bootstrap uncertainty ≤0.1\leq 0.1 ↑574642405647
MOFs with bootstrap uncertainty ≤0.2\leq 0.2 ↑661669594668

For KHRFK_H^{\mathrm{RF}}, the number of MOFs with Emin⁡ref−Emin⁡opt≤kBTE_{\min}^{\mathrm{ref}}-E_{\min}^{\mathrm{opt}}\leq k_BT increased from 637 to 672, where the two energies denote the lowest reference-sample energy and the lowest energy found by geometry optimization. The updated references improve both bootstrap precision and coverage of low-energy configurations.

Q3. How does the train–test split ensure that evaluation requires meaningful extrapolation to unseen MOFs?

The original evaluation uses a random in-distribution split, with 6,052 MOFs for training and 672 for testing and no shared CSD identifiers. To assess extrapolation beyond this setting, we additionally excluded selected metals, linkers, or topologies from training and retrained with the same protocol. We selected these groups from the most frequent categories in our dataset and evaluated on the corresponding original test MOFs.

CategoryGroup excluded from trainingTest MOFsMean ∣Δln⁡KH∣\lvert\Delta\ln K_H\rvert ↓Mean nESS (%) ↑
MetalCu1030.027 → 0.02659.5 → 51.3
MetalCd590.023 → 0.02455.7 → 50.4
MetalCo800.022 → 0.02848.2 → 40.9
Linker4,4′-bipyridine330.025 → 0.02568.3 → 66.8
LinkerTerephthalate (BDC)200.027 → 0.02763.0 → 62.0
Topologysql760.024 → 0.02557.2 → 54.2
Topologyhcb300.044 → 0.04849.6 → 48.6
Topologypcu170.018 → 0.01764.0 → 61.0

Arrows compare the model trained on all groups with the model retrained after excluding the specified group. Performance is largely maintained for unseen linkers and topologies, with little change in error and only small reductions in sampling efficiency. Excluding metals is more challenging because it requires extrapolation to elements absent from the proposal’s training data, going beyond the linker, topology, and size extrapolation. Even in this setting, the model retains useful sampling efficiency.

Size OOD evaluation requires reference data for additional MOFs beyond the original size range. We therefore randomly selected 200 MOFs with 201–300 atoms per unit cell using the same pore-accessibility criteria and generated new references. The model trained on MOFs with at most 200 atoms was evaluated without retraining.

Test setAtoms per unit cellTest MOFsMean ∣Δln⁡KH∣\lvert\Delta\ln K_H\rvert ↓Mean nESS (%) ↑
Original test set≤2006680.03152.0
Size OOD201–3002000.03561.7

The model maintains strong accuracy on the size OOD test set while achieving higher mean sampling efficiency.

AuthorsOpenReview ↗

Response to Reviewer DpT9 (3/3)

Q4. How does cross-MOF generalization compare with related work, and how challenging is the demonstrated extrapolation?

To our knowledge, few neural sampling studies evaluate cross-system generalization, and most examples concern molecular systems. Representative transferable BGs use classical force fields [1, 2]. For practical materials screening, we sought transfer across systems, a Boltzmann target defined by a DFT-trained MLIP, and lower actual GPU cost than conventional sampling (not number of evaluation), all within one method. Our method achieves all three, transferring to 672 held-out MOFs without per-material retraining. On the reference-qualified test sets, 1,024 draws yield lower GPU inference cost and lower mean absolute log-Henry error than the corresponding conventional baselines in every adsorption-strength quartile (Table 2).

Although not Boltzmann generators, AdsorbDiff [3] and AdsorbFlow [4] study cross-system generalization for adsorbate placement on catalyst surfaces, represented by two-dimensionally periodic slabs. They share our challenge of generating useful adsorption configurations for unseen material–adsorbate systems. However, their objective is to find configurations that relax to low-energy structures. Our task additionally requires estimating a configurational integral over a three-dimensional MOF pore network, accounting for both the energies and configurational volumes of thermodynamically important adsorption basins.

Around the same time as our submission, AdsorbSample [5] formulated catalyst-surface placement as conditional Boltzmann sampling under an MLIP energy. These developments motivate testing our architecture on the same placement benchmark, where a broader set of established baselines allows a direct assessment of its generation quality, cross-system transfer, and computational cost. This comparison complements our MOF experiments by showing how the architecture performs on a more widely studied adsorption task.

We trained our model with minimal necessary architectural changes, using OC20 IS2RE [6] for pretraining and the existing local-minimum configurations in OC20-Dense [7] for MLE fine-tuning, without generating additional SMC training data. On OC20-Dense ID and OOD subsets containing 41 and 50 systems, respectively, our model achieved the highest Pass@10 among the evaluated methods and the lowest estimated GPU cost in the OOD timing comparison.

MethodID Pass@10 (%) ↑OOD Pass@10 (%) ↑Estimated GPU cost (s/system) ↓
AdsorbML (heuristic)62.759.725.9
AdsorbDiff (retrained)34.136.044.8
AdsorbFlow (5 ODE steps)63.449.526.2
AdsorbFlow (30 ODE steps)68.358.239.2
Ours71.267.722.0

Pass@10 measures the probability that at least one of ten candidates is valid and within 0.1 eV of the reference energy. All methods use clean-slab inputs and UMA (MLIP) relaxation without bond restraints. Costs are estimated for ten candidates per OOD system on an H200, including generation and relaxation but excluding UMA feature precomputation. Estimates combine timing sources, with retrained AdsorbDiff using the released model’s generation timing.

GPU cost is particularly important for this task because proposal generation adds to the cost of subsequent relaxation. If this overhead outweighs the benefit of better proposals, random or heuristic search, as used in AdsorbML, can be more practical. We therefore compare the combined cost of generation and relaxation. We reproduced AdsorbFlow because its code was unavailable when these experiments were conducted. We do not directly compare with AdsorbSample because it uses bond restraints during geometry optimization and reports baseline results under that different protocol.

[1] Tan, C. B. et al. Amortized Sampling with Transferable Normalizing Flows. NeurIPS, 2025.

[2] Rehman, D. et al. Autoregressive Boltzmann Generators. ICML, 2026.

[3] Kolluru, A. and Kitchin, J. R. AdsorbDiff: Adsorbate Placement via Conditional Denoising Diffusion. ICML, 2024.

[4] Qiu, J. et al. AdsorbFlow: Energy-Conditioned Flow Matching Enables Fast and Realistic Adsorbate Placement. arXiv:2602.19289, 2026.

[5] Ram, A. et al. Diffusion Sampling of Adsorbate Configurations on Catalyst Surfaces. ICML AI for Science Workshop, 2026.

[6] Chanussot, L. et al. The Open Catalyst 2020 (OC20) Dataset and Community Challenges. ACS Catalysis, 2021.

[7] Lan, J. et al. AdsorbML: A Leap in Efficiency for Adsorption Energy Calculations Using Generalizable Machine Learning Potentials. npj Computational Materials, 2023.

Dear Reviewer DpT9,

Thank you again for your thoughtful review. We recently posted our response and additional experimental results. With the author-reviewer discussion period closing today (~ 13hr left), we would appreciate knowing whether these have addressed your concerns, or whether any points still need clarification.

Thank you

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.