acceptodds
Under review as a conference paper at ICLR 2027

MADField: Multi-fidelity Amortized Density Field for Adsorption in Nanoporous Materials

Abstract

High-throughput computational screening of nanoporous materials for gas storage and separation requires fast and accurate characterization of adsorption equilibrium. Particle-based grand canonical Monte Carlo (GCMC) and density-based classical density functional theory (cDFT) provide simulation-based estimates of gas uptake and adsorbate density fields, but their speed-accuracy tradeoff remains insufficient for large-scale screening. In this work, we address this gap with Multi-fidelity Amortized Density Field (MADField), which reframes adsorption prediction as equilibrium density-field estimation. MADField learns from two complementary fidelities, combining broad and scalable cDFT density supervision with GCMC density labels, and recovers gas uptake by integrating the predicted density field. MADField improves uptake accuracy over the strongest baselines by 5.7× for cDFT and 3.35× for GCMC, and its predicted fields accelerate cDFT solvers with a 5.7× reduction in walltime while recovering 37.78% of cases that fail under standard settings. We also evaluate MADField under realistic screening scenarios for CH4 working-capacity and low-pressure CO2 uptake screening, where it reduces the number of required GCMC candidate evaluations to obtain the top 0.1% pool by 8.0 and 11.5 times, respectively.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI/ML for physical sciences, Metal-organic Framework(MOF), adsorption surrogate model, material screening

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_D9VYOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MADField predicts physically scaled equilibrium adsorbate density fields with a Boltzmann-residual 3D Swin U-Net, pre-trained on ~566k PC-SAFT cDFT fields and LoRA-adapted to GCMC density labels. Uptake is obtained by integration. The predicted fields are also used to warm-start cDFT, recover failed labels, and screen ARC-MOF (CH₄ working capacity) and MOSAEC-DB (CO₂ at 0.15 bar).

Critical strengths

The dataset and solver-in-the-loop pipeline are useful; recovered labels must be solver-verified rather than taken from network outputs. The work has a clear physical motivation. The controlled CH₄/Xe model includes framework-bootstrap CIs and fine-tuning seeds (Table 18). These experiments are fairly credible.

Critical weaknesses
    1. The central control is confounded. The key evidence for “density supervision beats scalar regression” is Table 14 (7.47 vs 1.31) and Table 15 (7.531 vs 1.996). Per lines 1350–1352, the scalar head mean-pools the final tokens, replaces the voxel decoder, and drops the Boltzmann-residual parameterization. This changes the readout, decoder/physics prior, and supervision at once. The density-supervised “w/o Boltzmann residual” variant reaches 1.55, close to 1.31, so the main confound is mean-pool readout without a decoder versus density supervision. The authors acknowledge (lines 1357, 1406) that the variants compare complete target/head schemes.

Treating uptake as a spatial integral of local pore contributions is itself a strong inductive bias, independent of whether density labels are used. Please add an integral-readout, scalar-supervised control on the same backbone and decoder. If density labels still add a substantial gain over this control, the central claim holds and I would raise my score to 3; otherwise, the title and abstract should reposition the contribution around the dataset and multi-fidelity transfer.

On related work: SpatialRead (Zou et al., ICLR 2026, “From atom to space: a region-based readout function for spatial properties of materials”) also formulates adsorption as a spatial integral of local contributions learned from scalar labels. This overlaps with the integral-readout part of MADField and reduces its novelty; the distinct contribution then rests on density supervision, which is what the control in point 1 should isolate. A comparison with SpatialRead, or the equivalent integral-readout control, would separate the two.

  1. The CO₂ results do not support the thesis and use privileged inputs. The CO₂ extension ends with a scalar amplitude head (exp(2tanh(·)), range ≈0.14–7.4×) trained mainly on an uptake loss; an ablation is needed to show how much of the CO₂ accuracy comes from this head. The CO₂ density itself is poor (relative L1 61.8%; Tanimoto 0.634). Only MADField receives REPEAT charges and electrostatic channels (Table 12 caption). This differs from the setting of the other compared methods and may confound the contribution of the method with that of the extra information.

  2. The screening results lack the model pretrained only on cDFT. cDFT ranks single-pressure CH₄ uptake with Spearman 0.999 against GCMC at 27 s/state (Tables 24–25). I think cDFT is still much slower than a neural network, so it may not be a suitable baseline, but I am curious about the results of the model pretrained only on cDFT. Recall–budget curves are needed for MADField-cDFT without GCMC adaptation.

  3. Uncertainty remains unclear. Only the controlled CH₄/Xe model has intervals. The main tables, PRAM transfer (138 states; the “AC” family has 3 states yet carries 1/4 of the mean weight) and the complete-recovery budgets need intervals. That said, the improvements in the tables are fairly clear, so omitting intervals is largely reasonable, but confidence intervals would make the results more rigorous. The multi-fidelity gain is 7.2% at full labels and 41% at 10% labels; the abstract should state it this way.

Minor comments

    • The main tables report no run-to-run variance for the baselines or MADField. The margins are large enough not to change the conclusions, but please run them several times and report means and standard deviations.
  • Table 3's GCMC mean mixes a 967-framework CH₄/Xe test with a 90-MOF CO₂ test.
  • The gate is called h_θ in Eq. 7 but g_θ in Eq. 20.
  • The “Mat. ✓” entry in Table 1 should say zero-shot.

The dataset and solver-in-the-loop pipeline are valuable, but the key claim that density supervision improves uptake prediction is not yet supported by an unconfounded control. I would raise my score to 3 if the integral-readout, scalar-supervised control clearly shows a substantial gain from density labels and the CO₂ amplitude-head ablation is added. The overlap with SpatialRead also lowers my assessment of novelty somewhat, and this is reflected in the score.

Questions for authors

(1) What is the error of the integral-readout, scalar-supervised control? (2) What is the CO₂ MAE without the amplitude head, and with charge inputs given to the baselines? (3) What are the recall curves of MADField-cDFT without GCMC adaptation? (4) What are the seed or bootstrap intervals for the main tables and the screening budgets?

6 comments

We thank the reviewer for the careful and constructive review. We ran the requested integral-readout control, the CO2\mathsf{CO_2} amplitude-head ablation, charge-aware baselines, and the MADField-cDFT screening comparison, and added confidence intervals. The results are given below and in the revised manuscript, where changes are marked in red.

W1 / Q1. Isolating the effect of density supervision

We agree that the original scalar-head comparison changed both the supervision target and the readout. We therefore trained the requested control, “Scalar integral” in Table R1, which uses the same backbone, decoder, Boltzmann-residual parametrization, and integral readout as our model, “Density MADField.” The only difference is the supervision: Scalar integral is trained solely on scalar uptake labels, whereas Density MADField additionally receives density-field supervision. Under this controlled comparison, density supervision reduces uptake MAE by 37.0–60.2% relative to scalar uptake supervision.

Table R1 (Appendix I.8, Table 26). Controlled scalar-supervision ablation. Uptake MAE (↓) is reported in cm3(STP)/gcm^3(\mathrm{STP})/g. The MAE reduction is computed relative to Scalar integral, with 95% confidence intervals obtained by paired bootstrap over matched test samples. The CO2\mathsf{CO_2} column follows the training protocol of the submitted model and reports mean ± SD over two seeds.

cDFT only (9 adsorbates)GCMC only (CH4\mathsf{CH_4}, Xe)cDFT → GCMC (CH4\mathsf{CH_4}, Xe)cDFT → GCMC (CO2\mathsf{CO_2})
Scalar integral3.1064.7195.10113.554±0.21813.554 \pm 0.218
Density MADField1.3462.2302.0288.542±0.3288.542 \pm 0.328
MAE reduction (95% CI)56.6% (50.9–61.9%)52.7% (50.0–55.4%)60.2% (57.6–62.8%)37.0% (23.6–48.4%)

Relation to SpatialRead. The integral readout is not our claimed contribution. SpatialRead (Zou et al., ICLR 2026) shows that spatially integrating learned local contributions can itself provide an effective inductive bias for adsorption-property prediction from scalar labels. Our new Scalar integral control therefore directly tests whether this inductive bias alone explains MADField’s gain. Table R1 shows that, even when the spatial-integral architecture is held fixed, direct density supervision still reduces uptake MAE by 37.0–60.2%.

Our contribution instead centers on learning the physical, unnormalized adsorption density as a common prediction target across simulation fidelities. This enables abundant, scalable cDFT density labels to pre-train the model before adaptation to much more expensive GCMC density labels, reducing the amount of high-fidelity density supervision required. The predicted field also determines uptake by integration and directly supports spatial adsorption analysis and cDFT solver initialization. We now cite and discuss SpatialRead in Section 2.2 and Appendix I.8.

W2 / Q2. CO2\mathsf{CO_2}: density supervision, amplitude head, and charge inputs

To address this concern, we ran three controlled experiments isolating density supervision, the CO2\mathsf{CO_2} amplitude head, and framework-charge inputs. The results show that MADField's high CO2\mathsf{CO_2} accuracy compared to other baselines is not explained by the amplitude head or unequal access to charge information, while density supervision retains a substantial benefit under matched inputs and architecture. Specifically, (i) with identical charge inputs and architecture, density supervision reduces CO2\mathsf{CO_2} MAE by 37.0% (Table R1), (ii) removing the amplitude head leaves accuracy unchanged (Table R2), and (iii) scalar baselines given the same charge information stay at 20.99 MAE or above, compared with 8.76 for MADField (Table R3).

(i) Density supervision under matched inputs. The scalar-integral control and MADField receive the same charge inputs and have the same architecture, including the amplitude head. Density supervision reduces CO2\mathsf{CO_2} MAE from 13.554 to 8.542 (37.0%, last column of Table R1). Although CO2\mathsf{CO_2} remains the most challenging density-prediction task (relative L1 61.8%), the matched control shows that density supervision still substantially improves uptake prediction.

(ii) Amplitude head. We retrained MADField without the global amplitude head that rescales the predicted density field.

Table R2 (Appendix I.9). CO2\mathsf{CO_2} uptake MAE (↓) in cm3(STP)/gcm^3(\mathrm{STP})/g.

With amplitude headWithout amplitude head
MADField, retrained (2 seeds)8.542 ± 0.3288.348 ± 0.299

Removing the amplitude head does not degrade CO2\mathsf{CO_2} uptake accuracy (paired 95% CI of the difference [−1.297, 0.773]).

(iii) Charge inputs for the baselines. CO2\mathsf{CO_2} is a quadrupolar adsorbate whose uptake depends on electrostatic interactions with the framework. The GCMC reference labels are themselves generated with REPEAT framework charges and a charged CO2\mathsf{CO_2} model. To directly address the concern that MADField may benefit from privileged electrostatic information, we matched the charge information available to the scalar baselines (RetNet, Uni-MOF, and MOFTransformer) in two ways (details in Appendix I.9):

  • Charge adapter. Per-atom charges are embedded and added to the atom embeddings of Uni-MOF and MOFTransformer. For RetNet, which takes voxel grids, the electrostatic potential computed from the charges is added after its first convolutional layer.
  • Electrostatic field. The electrostatic interaction between the framework charges and a CO2\mathsf{CO_2} molecule is computed on a grid, encoded, and given to each model as an additional input.

The added modules have about 12.47M parameters, matching MADField's charge-related modules. Each "without charge" control has identical modules with the charge inputs set to zero.

Table R3 (Appendix I.9, Table 27). CO2\mathsf{CO_2} uptake MAE (↓) in cm3(STP)/gcm^3(\mathrm{STP})/g, mean ± SD over three seeds. Charge gain is defined as MAE(without charge) − MAE(with charge).

ModelHow charges are givenWithout chargeWith chargeCharge gain, paired 95% CI
RetNetCharge adapter28.003 ± 0.46221.376 ± 0.312[2.809, 11.090]
RetNetElectrostatic field28.058 ± 0.43121.466 ± 0.374[2.446, 11.299]
Uni-MOFCharge adapter23.878 ± 0.38523.414 ± 0.632[−0.738, 1.574]
Uni-MOFElectrostatic field22.343 ± 0.16423.305 ± 0.961[−1.603, −0.334]
MOFTransformerCharge adapter22.972 ± 0.22022.748 ± 0.228[−0.193, 0.673]
MOFTransformerElectrostatic field21.616 ± 0.20320.992 ± 0.511[−0.126, 1.486]

Charge-aware retraining improves RetNet, while producing small or inconsistent changes for Uni-MOF and MOFTransformer. Importantly, the best charge-aware scalar baseline still has an MAE of 20.992, compared with 8.763 for MADField.

W3 / Q3. Screening with MADField-cDFT (no GCMC adaptation)

To directly assess how much of the screening performance is obtained from cDFT pretraining alone, we added MADField-cDFT to the recall-budget curves in Appendix I.12, Figure 11. Here, MADField-cDFT denotes the model pretrained only on cDFT densities, while MADField-GCMC denotes the model after GCMC adaptation. The cDFT-only model provides strong CH4\mathsf{CH_4} ranking performance, whereas CO2\mathsf{CO_2} screening improves substantially after GCMC adaptation.

Table R4 (Table 31). CH4\mathsf{CH_4} working-capacity screening on the original 254,198-candidate pool: number of top-ranked candidates that must be evaluated to recover each fraction of the 166 targets with working capacity ≥200 cm3(STP)/cm3cm^3(\mathrm{STP})/cm^3 (↓).

Ranking method60% recall80% recall100% recall
MADField-cDFT1,3902,20611,666
MADField-GCMC17572316,600

MADField-GCMC requires substantially fewer screened candidates at 60–80% recall than MADField-cDFT. At 100% recall, MADField-cDFT requires fewer candidates, reflecting a small number of tail cases that are ranked lower after GCMC adaptation. Overall, however, MADField-GCMC provides the better recall–budget trade-off across the screening curve.

Table R5 (Table 32). CO2\mathsf{CO_2} screening at 0.15 bar, same quantity as Table R4. The pool is all 17,152 frameworks held out from CO2\mathsf{CO_2} training, with 18 targets (true top 0.1%). The paper’s seven-model comparison uses the 15,572 of these that every baseline can process, with 16 targets.

Ranking method60% recall80% recall100% recall
MADField-cDFT15,44516,45916,737
MADField-GCMC1529166

For CH4\mathsf{CH_4}, MADField-cDFT retains much of the screening performance, indicating that the cDFT-pretrained density model captures a ranking signal that transfers well to the GCMC target. In contrast, the CO2\mathsf{CO_2} ranking improves substantially after GCMC adaptation. This difference is consistent with the lower cDFT–GCMC agreement for CO2\mathsf{CO_2} than for CH4\mathsf{CH_4} observed in Table 25.

Notably, the benefit of cDFT pretraining is not limited to reproducing cDFT rankings. In the controlled CH4\mathsf{CH_4}/Xe experiment (Table 17), its uptake-MAE reduction grows from 7.2% with all GCMC labels to 41.0% when only 10% of GCMC labels are available.

Reviewer_5kUvOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents MADField, a multi-fidelity surrogate for equilibrium gas adsorption in nanoporous materials. A periodic 3D Swin model predicts positive density fields using a Boltzmann-residual parameterization and obtains uptake by integration. The model is pretrained on cDFT fields and adapted to GCMC fields, with additional charge-aware modules for CO₂. The predicted fields also support material screening and cDFT solver initialization.

Critical strengths
  1. The formulation connects spatial adsorption structure, uptake prediction, and solver initialization through a common absolute-density representation.
  2. The matched-backbone experiments support the density-supervised formulation.
  3. The screening experiments demonstrate practical value at substantial scale.
Critical weaknesses
  1. The CO₂ comparison confounds method design with additional physical inputs and capacity. Only MADField receives REPEAT charges and derived charge grids (Table 12 and Appendix G), while its extension increases the parameter count from 6.35M to 21.36M (Appendix D.2). The contributions of density supervision, charge information, and added capacity therefore remain insufficiently separated.
  2. The uptake comparison omits a directly relevant density baseline. The GCMC-adapted SorbIIT predicts absolute densities under fixed target scaling and requires no reference uptake at inference (Appendix G). However, Table 3 excludes its integrated uptake predictions. This comparison is needed to establish MADField’s advantage over the adapted absolute-density alternative.
  3. Computational efficiency is not evaluated under a matched label-generation budget. Under the protocols in Table 24, a GCMC density label costs 18,468 seconds versus 118.3 seconds for a scalar label, approximately 156 times more. The label-count comparison leaves open whether the advantage persists when scalar models receive more labels for the same simulation budget, and whether screening savings offset the initial data-generation and training costs.
Questions for authors
  1. How much of the CO₂ improvement remains when the strongest scalar baseline receives comparable charge information, model capacity, and tuning resources? Which charge-aware components contribute most?
  2. What uptake MAE and screening performance result from integrating the GCMC-adapted SorbIIT’s absolute-density predictions?
  3. Does density supervision retain its advantage under equal simulation budgets? At what screening-library size do the savings offset the additional data-generation and training costs?
5 comments

We thank the reviewer for the constructive feedback. To address the concerns, we ran a CO2\mathsf{CO_2} control matched in charge inputs and capacity, trained charge-aware scalar baselines, evaluated SorbIIT's integrated uptake, and added a cost analysis. The results are given below and in the revised manuscript, where changes are marked in red.

W1. CO2\mathsf{CO_2} inputs, model capacity, and amplitude correction

To separate density supervision, charge information, and model capacity, we ran three controlled experiments. With charge inputs, capacity, and amplitude head matched, density supervision reduces CO2\mathsf{CO_2} MAE by 37.0%. Scalar baselines given the same charge information and matched capacity stay at 20.99 MAE or above, compared with 8.76 for MADField.

(i) Density supervision under matched inputs and capacity. "Scalar integral" uses the same backbone, decoder, Boltzmann-residual parametrization, and integral readout as our model, "Density MADField." The only difference is the supervision: Scalar integral is trained solely on scalar uptake labels, whereas Density MADField additionally receives density-field supervision. For CO2\mathsf{CO_2}, both models include the charge modules and the global amplitude head that rescales the predicted density field, and both have exactly 21,359,870 parameters (the 21.36M CO2\mathsf{CO_2} model of Appendix D.2).

Table R1 (Appendix I.8, Table 26). Controlled scalar-supervision ablation. Uptake MAE (↓) is reported in cm3(STP)/gcm^3(\mathrm{STP})/g. The MAE reduction is computed relative to Scalar integral, with 95% confidence intervals obtained by paired bootstrap over matched test samples. The CO2\mathsf{CO_2} column follows the training protocol of the submitted model and reports mean ± SD over two seeds.

cDFT only (9 adsorbates)GCMC only (CH4\mathsf{CH_4}, Xe)cDFT → GCMC (CH4\mathsf{CH_4}, Xe)cDFT → GCMC (CO2\mathsf{CO_2})
Scalar integral3.1064.7195.10113.554±0.21813.554 \pm 0.218
Density MADField1.3462.2302.0288.542±0.3288.542 \pm 0.328
MAE reduction (95% CI)56.6% (50.9–61.9%)52.7% (50.0–55.4%)60.2% (57.6–62.8%)37.0% (23.6–48.4%)

Density supervision reduces CO2\mathsf{CO_2} MAE from 13.554 to 8.542. The same control gives 52.7–60.2% reductions for CH4\mathsf{CH_4}/Xe, where no charge inputs are involved.

(ii) Charge inputs for the baselines. CO2\mathsf{CO_2} is a quadrupolar adsorbate whose uptake depends on electrostatic interactions with the framework. The GCMC reference labels are themselves generated with REPEAT framework charges and a charged CO2\mathsf{CO_2} model. To address the concern that MADField benefits from additional physical inputs, we matched the charge information available to the scalar baselines (RetNet, Uni-MOF, and MOFTransformer) in two ways (details in Appendix I.9):

  • Charge adapter. Per-atom charges are embedded and added to the atom embeddings of Uni-MOF and MOFTransformer. For RetNet, which takes voxel grids, the electrostatic potential computed from the charges is added after its first convolutional layer.
  • Electrostatic field. The electrostatic interaction between the framework charges and a CO2\mathsf{CO_2} molecule is computed on a grid, encoded, and given to each model as an additional input.

The added modules have about 12.47M parameters, matching MADField's charge-related modules. Each "without charge" control has identical modules with the charge inputs set to zero.

Table R2 (Appendix I.9, Table 27). CO2\mathsf{CO_2} uptake MAE (↓) in cm3(STP)/gcm^3(\mathrm{STP})/g, mean ± SD over three seeds. Charge gain is defined as MAE(without charge) − MAE(with charge).

ModelHow charges are givenWithout chargeWith chargeCharge gain, paired 95% CI
RetNetCharge adapter28.003 ± 0.46221.376 ± 0.312[2.809, 11.090]
RetNetElectrostatic field28.058 ± 0.43121.466 ± 0.374[2.446, 11.299]
Uni-MOFCharge adapter23.878 ± 0.38523.414 ± 0.632[−0.738, 1.574]
Uni-MOFElectrostatic field22.343 ± 0.16423.305 ± 0.961[−1.603, −0.334]
MOFTransformerCharge adapter22.972 ± 0.22022.748 ± 0.228[−0.193, 0.673]
MOFTransformerElectrostatic field21.616 ± 0.20320.992 ± 0.511[−0.126, 1.486]

The best charge-aware scalar baseline still has an MAE of 20.992, compared with 8.763 for MADField.

(iii) Amplitude head. We retrained MADField without the amplitude head.

Table R3 (Appendix I.9). CO2\mathsf{CO_2} uptake MAE (↓) in cm3(STP)/gcm^3(\mathrm{STP})/g.

With amplitude headWithout amplitude head
MADField, retrained (2 seeds)8.542 ± 0.3288.348 ± 0.299

Removing the amplitude head does not degrade CO2\mathsf{CO_2} uptake accuracy (paired 95% CI of the difference [−1.297, 0.773]).

W2. SorbIIT integrated uptake

As suggested, we integrated the predicted densities of the GCMC-adapted SorbIIT, which predicts absolute density without reference uptake at inference (Appendix G), using its existing checkpoints. MADField's uptake MAE is 5.4–31× lower.

Table R4 (Appendix I.13, Table 34). Uptake MAE (↓) in cm3(STP)/gcm^3(\mathrm{STP})/g.

GasSorbIITMADField
CH4\mathsf{CH_4}31.211.41
Xe70.662.25
CO2\mathsf{CO_2}47.238.76

Screening. On the 17,152 frameworks held out from CO2\mathsf{CO_2} training, with 18 targets (true top 0.1%), two SorbIIT checkpoints need 9,557 and 10,406 evaluations to recover all targets, compared with 166 for MADField. For CH4\mathsf{CH_4} working capacity, the SorbIIT checkpoints predict only at their trained pressures (1 and 65 bar), while the task also requires 5.8 bar. MADField predicts at 5.8 bar without GCMC supervision at that pressure.

W3. Computational cost and equal simulation budget

We agree that GCMC density labels are more expensive than scalar labels: 18,468 s versus 118.3 s per label, while a cDFT density label takes 27.1 s (Table 24). To address when screening savings offset this cost, we added a break-even analysis (Appendix I.15). If all density labels were generated from scratch, the cost would be recovered after about 8.7 million screened candidates, or about 34 ARC-MOF-scale screens.

Break-even. In CH4\mathsf{CH_4} working-capacity screening, MADField reduces the candidates that need follow-up GCMC from 52.16% (RetNet, the best baseline) to 6.53%. This saves about 7,600 worker-hours of GCMC per screen of the 254K-candidate ARC-MOF pool (two pressures per candidate). Generating our full GCMC density training set and cDFT pre-training set from scratch would cost about 262,000 worker-hours more than the scalar labels a baseline needs.

Three points reduce this cost in practice:

  1. One-time data cost. Training data is generated once and reused across screening tasks. Each later screen needs only neural inference and GCMC verification of the shortlisted candidates. This extends to new operating conditions: MADField predicts CH4\mathsf{CH_4} uptake at 5.8 bar without GCMC supervision at that pressure.
  2. Reused CH4\mathsf{CH_4}/Xe labels. We used the published CH4\mathsf{CH_4}/Xe density data of Burner et al., so the CH4\mathsf{CH_4} screening required no new GCMC density-label generation.
  3. Transfer to large disordered systems. Without retraining, MADField transfers to PRAM, a set of 138 states of non-MOF porous polymers (PIM, HCP), kerogens, and amorphous carbons (AC) (Table 4).

Table R5 (paper Table 4). GCMC uptake MAE (↓) in cm3(STP)/gcm^3(\mathrm{STP})/g on PRAM.

ModelPIMHCPKerogenACMean
RetNet (best baseline)2.281.946.369.515.02
MADField1.721.212.111.451.62

These amorphous structures require much larger simulation cells than a MOF unit cell, so GCMC on them is considerably more expensive, whereas MADField's inference cost is the same as for a MOF. The savings per avoided GCMC run are therefore larger for such systems.

Reviewer_Sp9EOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper predicts adsorbate density fields instead of scalar uptake, then integrates the field to get uptake. The model is pre-trained on lots of cDFT data and fine-tuned on a small set of GCMC labels. It also helps in practice: predicted fields speed up cDFT solves, and predicted uptakes cut down the GCMC runs needed for screening. The idea is neat and the dataset is a real contribution.

Critical strengths
  • using density as the shared target across fidelities makes physical sense, and you can actually look at what the model predicts.
  • the paper closes the loop: faster cDFT, recovered failed runs, cheaper screening.
  • the 566K cDFT fields would be a useful public dataset.
  • splits and ablations are handled carefully.
Critical weaknesses
  • Pre-training doesn't buy much at full data. It gives about 7% (Table 17). Full fine-tuning also beats the rank-4 adapter you actually use. So where does the 3.35× come from?
  • The baselines may be data-starved. If only MADField sees the 566K cDFT samples, the headline ratios mostly reflect data access. Pre-train the baselines on cDFT too. The main results also need more seeds and CIs.
  • The evaluation protocol changes between tables (masked integration with BF16 vs. full domain with FP32).
  • The "density beats scalar" result isn't clean. The GCMC density labels are normalized to the true uptake, so they already contain the answer. The compared variants also differ in training budget, precision and losses. A shape-only ablation would settle it.
Questions for authors

please reference above

5 comments

We thank the reviewer for the constructive feedback. To address the concerns, we trained matched scalar-supervised controls, pre-trained all five external baselines on cDFT, and re-evaluated our checkpoints under both numerical precisions. The results are given below and in the revised manuscript, where changes are marked in red.

W1. Attribution of the 3.35× improvement

The 3.35× improvement compares the complete MADField pipeline with the strongest external scalar baseline, so it combines several factors. To separate them, we ran a controlled scalar-supervision ablation (Table R1) and revisited the label-efficiency results (Table R2). Density supervision accounts for most of the gain, while cDFT pre-training mainly improves data efficiency. On CH4\mathsf{CH_4}/Xe, starting from the strongest retrained external baseline (RetNet, 9.348, Table R3), our architecture trained on scalar labels reaches 4.719, adding density supervision reaches 2.230, and adding cDFT pre-training reaches 2.028.

Density supervision. "Scalar integral" uses the same backbone, decoder, Boltzmann-residual parametrization, and integral readout as our model, "Density MADField." The only difference is the supervision: Scalar integral is trained solely on scalar uptake labels, whereas Density MADField additionally receives density-field supervision. Under this controlled comparison, density supervision reduces uptake MAE by 37.0–60.2%.

Table R1 (Appendix I.8, Table 26). Controlled scalar-supervision ablation. Uptake MAE (↓) is reported in cm3(STP)/gcm^3(\mathrm{STP})/g. The MAE reduction is computed relative to Scalar integral, with 95% confidence intervals obtained by paired bootstrap over matched test samples. The CO2\mathsf{CO_2} column follows the training protocol of the submitted model and reports mean ± SD over two seeds.

cDFT only (9 adsorbates)GCMC only (CH4\mathsf{CH_4}, Xe)cDFT → GCMC (CH4\mathsf{CH_4}, Xe)cDFT → GCMC (CO2\mathsf{CO_2})
Scalar integral3.1064.7195.10113.554±0.21813.554 \pm 0.218
Density MADField1.3462.2302.0288.542±0.3288.542 \pm 0.328
MAE reduction (95% CI)56.6% (50.9–61.9%)52.7% (50.0–55.4%)60.2% (57.6–62.8%)37.0% (23.6–48.4%)

cDFT pre-training. Its benefit grows as GCMC labels become scarce, from 7.2% with all labels to 41.0% with 10% of the labels.

Table R2 (paper Table 17). Uptake MAE (↓) in cm3(STP)/gcm^3(\mathrm{STP})/g of density-supervised models versus the fraction of GCMC labels used. Evaluated in bfloat16 (BF16), as in the paper.

GCMC labels usedGCMC onlycDFT → GCMCMAE reduction
10%4.7152.78241.0%
20%3.2972.29230.5%
50%2.4232.06814.7%
100%2.1521.9967.2%

LoRA versus full fine-tuning. Full fine-tuning is slightly more accurate than the rank-4 LoRA adapter (1.930 vs. 1.996), but it trains 6.354M instead of 1.573M parameters. LoRA keeps the cDFT backbone frozen, so a single backbone serves both fidelities.

W2. cDFT pretraining for external baselines, seeds, and CIs

To address the concern about data access, we pre-trained all five external scalar baselines on cDFT uptake labels and then fine-tuned them on GCMC, with the same training budget and test set as their GCMC-only counterparts. cDFT pre-training leaves their accuracy essentially unchanged, and all five remain above 9.3 MAE, compared with 2.028 for MADField.

Table R3 (Appendix I.10, Table 28). CH4\mathsf{CH_4}/Xe uptake MAE (↓) in cm3(STP)/gcm^3(\mathrm{STP})/g, mean ± SD over three seeds. The baselines are newly retrained and reported as the average over CH4\mathsf{CH_4} and Xe, whereas Table 3 of the paper reports the original checkpoints per gas.

BaselineGCMC onlycDFT → GCMC
Uni-MOF9.817 ± 0.0759.923 ± 0.089
RetNet9.348 ± 0.0879.356 ± 0.122
MOFTransformer11.631 ± 0.25311.673 ± 0.147
DeepSorption13.039 ± 0.37812.924 ± 0.812
IsothermNet13.689 ± 0.85212.750 ± 0.331

The gap to MADField is significant for all five baselines (paired bootstrap 95% CIs exclude zero). Scalar cDFT pre-training also does not help our own Scalar integral control (5.101 vs. 4.719, Table R1), whereas density pre-training reduces MAE by up to 41.0% (Table R2). This suggests that cDFT pre-training is most useful when the density field is the shared target.

CIs and seeds. We added 95% bootstrap CIs for the main MADField results: 1.346 [1.205, 1.513] for cDFT and 2.028 [1.895, 2.162] for GCMC CH4\mathsf{CH_4}/Xe (Appendix I.11). Over three fine-tuning seeds, the CH4\mathsf{CH_4}/Xe model gives 2.018 ± 0.021 (Section I.2).

W3. Evaluation protocol

To check the sensitivity to the evaluation protocol, we re-evaluated the same checkpoints under both numerical precisions, bfloat16 (BF16) and 32-bit floating point (FP32), and under both integration supports. The differences are at most 0.04 MAE and do not change any comparison.

Table R4 (Appendix I.11). Uptake MAE in cm3(STP)/gcm^3(\mathrm{STP})/g of the same checkpoint under each precision.

ModelBF16FP32
MADField-cDFT1.3071.346
MADField-GCMC, CH4\mathsf{CH_4}/Xe1.9962.028

Integrating the prediction over the pore region or over the full grid changes MAE by only about 10−810^{-8} cm3(STP)/gcm^3(\mathrm{STP})/g. All new experiments in this response are evaluated in FP32, and Table R2 uses BF16 as in the paper. Appendix I.11 now states the precision and integration support.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.