acceptodds
Under review as a conference paper at ICLR 2027

Learning Hierarchical Representations of Metal–Organic Frameworks with Coupled Euclidean and Hyperbolic Geometry

Abstract

Learning representations of structured materials requires combining local geometry with connectivity across chemical building units. Metal–organic frameworks (MOFs) make this challenge explicit: their properties depend on both atomic environments and the organization of metal-containing units and organic linkers. We introduce , a graph representation model that couples Euclidean and hyperbolic geometry. A Euclidean branch encodes periodic local interactions, while a Lorentz branch processes an atomic covalent graph enriched with building-block features. Graph-level attention combines the two representations for property prediction. Building blocks augment atom features rather than define a separate coarse-grained graph. Across the reported hMOF and CoRE MOF adsorption benchmarks, achieves the lowest MAE among the evaluated models; additional comparisons cover MOF-specific architectures, repeated training runs, and accessible surface area. Geometric diagnostics show lower normalized hyperbolicity than degree-matched rewired controls and better graph-distance preservation for the evaluated Lorentz encoder. However, hyperbolicity has little association with per-structure prediction gains. These findings support coupled geometry as a useful inductive bias for MOF pore-related properties, while separating geometric motivation from claims about the causes of predictive improvement.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Geometric Deep Learning, Graph Representation Learning, Hyperbolic Neural Networks, Metal–Organic Frameworks

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_7CN5OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes GeoDuet, a dual-branch model combining Euclidean local geometry and Lorentz-based atomic connectivity for metal–organic framework (MOF) property prediction. The two graph-level representations are fused through attention. GeoDuet achieves strong performance on gas adsorption and accessible surface area prediction, supported by component ablations and geometric analyses.

Critical strengths

1. Effective Integration of Complementary Representations

The combination of periodic local geometry and building-block-aware atomic connectivity is well motivated and provides a coherent representation for MOFs.

2. Strong Empirical Performance

GeoDuet achieves the lowest reported MAE across the evaluated CH₄, CO₂, and Ar adsorption conditions. Additional surface-area experiments and geometric analyses further support the practical value of the approach.

Critical weaknesses

Insufficient Isolation of Hyperbolic Geometry's Contribution

As acknowledged by the authors, the contribution of hyperbolic geometry to predictive performance has not been sufficiently isolated. Table 6 shows that incorporating building-block features provides a substantially larger improvement than introducing the Lorentz branch. Moreover, the primary component ablation is restricted to CH₄ adsorption at 4.5 bar, whereas the main benchmarks cover multiple gases and pressure conditions. This makes it difficult to assess whether the contribution of the Lorentz branch generalizes across different prediction settings. Broader ablation studies, including a Lorentz-only configuration and a matched Euclidean–Euclidean baseline, would provide stronger evidence for the independent contribution of hyperbolic geometry.

Questions for authors

1. Comparison with Existing Hyperbolic GNNs

Have the authors considered comparing GeoDuet with representative hyperbolic GNN architectures under the same MOF prediction settings? Such comparisons would help clarify the advantages of the proposed approach.

2. Attention-Based Fusion

Given that attention is applied to only two graph-level representations, could the authors clarify its advantage over simpler fusion strategies, such as concatenation followed by an MLP?

3. Implementation Details

Could the authors provide more details on the message-passing functions (fEf_E, gEg_E, fLf_L, and gLg_L), the incorporation of geometric edge features, and the construction of building-block descriptors?

2 comments

Thank you for the suggestions. We added component comparisons for both CH4 and CO2, an existing hyperbolic-GNN reference, and attention-versus-MLP experiments. We provide the requested message-passing and feature-construction details below.

Component contributions across CH4 and CO2.

The new comparison includes the capacity-matched dual-Euclidean alternative. GeoDuet achieves the lowest test MAE among the evaluated variants on both adsorption tasks.

All models below were trained from scratch on the same 20,000 training structures, with 13,765 validation structures and 13,765 test structures, using seed 17. They have same layers. All models share a 50-epoch training schedule. Values are test MAE in mol/kg; BB denotes building-block features.

ModelCH4, 4.5 barCO2, 2.5 bar
Euclidean only0.55661.1692
Euclidean + BB0.39330.7530
Dual Euclidean + BB0.40500.7682
GeoDuet without BB0.56281.1899
GeoDuet0.38390.7403

The dual-Euclidean model retains the graph inputs, BB features, and fusion mechanism while replacing the Lorentz branch; its parameters are within 0.2% of GeoDuet's at this scale. The results show a GeoDuet advantage on both CH4 and CO2. Comparing Euclidean only with Euclidean + BB under the same reference schedule also shows the benefit of building-block information. We will extend Table 6 with these explicitly defined feature conditions and report the full-data Euclidean result of 0.17770 as an atomic-feature reference.

Q1. Comparison with an existing hyperbolic GNN.

GeoDuet achieves lower test MAE than the adapted HGCN on both tasks. We used the reference HGCN implementation, with atomic and BB features, four hyperboloid convolution layers, tangent-space pooling, and a regression head. Both models use the same 20,000/13,765/13,765 train/validation/test structures and seed 17.

ModelCH4, 4.5 barCO2, 2.5 bar
HGCN, standalone adaptation0.76501.6089
GeoDuet0.38390.7403

The results show that our method significantly outperforms HGCN, as we adapt hyperbolic networks to crystal material prediction tasks, enabling them to better capture the spatial structures of crystal materials.

Q2. Why attention over only two representations?

Attention learns input-dependent interactions between the Euclidean and Lorentz graph representations. With two representations, its role is to adapt their combination to each MOF. We tested this choice against concatenation followed by an MLP, including a parameter-matched MLP and a larger MLP.

The experiment trains fresh readouts on frozen full-data GeoDuet encoders with four layers and width 256, originally trained with attention fusion on 110,121 adsorption-labelled structures. Readout training uses 10,000 structures, with 13,765 validation and 13,765 test structures. The table reports mean ± sample standard deviation across three readout-initialization seeds, 17, 42, and 123. Each readout uses its stated training configuration; the configurations differ across readout types.

ReadoutTrainable parametersCH4, 4.5 barCO2, 2.5 bar
Concatenation + matched MLP329,3030.17126 ± 0.000120.33924 ± 0.00017
Concatenation + larger MLP460,0330.17034 ± 0.000240.33953 ± 0.00012
Two-token attention329,2170.16868 ± 0.000650.33898 ± 0.00036

Attention achieves the lowest mean MAE on both targets, with essentially the same parameter count as the matched MLP and fewer parameters than the larger MLP. The gain is more pronounced for CH4; the CO2 means are close. Holding the final CO2 attention configuration fixed and adding two initialization seeds gives 0.33892 ± 0.00030 over five seeds. These results support attention as a parameter-efficient fusion module on the learned GeoDuet representations; the reported variability concerns readout initialization with the encoders held fixed.

Q3. Message passing, geometry, and building-block descriptors.

For the Euclidean branch, let q_ij = [u_i; u_j; e_ij]. The message function is

f_E(u_i, u_j, e_ij) = sigmoid(BN(A_E(q_ij))) ⊙ B_E(q_ij),

where A_E and B_E are separate two-layer MLPs with SiLU activations. Messages are summed over neighbors; g_E adds the batch-normalized sum to u_i, followed by ReLU. Edge embeddings encode inverse distance, -0.75/r, using radial bases. Lattice-vector lengths and angles between edges and lattice vectors provide additional radial/spherical basis features for updating the edge embeddings.

For the Lorentz branch, neighboring features are combined by Lorentz concatenation. The message f_L is a channel-wise sigmoid gate multiplied by a nonlinear transformation of that concatenation. Both transformations use two Lorentz fully connected layers with intermediate Lorentz batch normalization and SiLU activation. The update g_L sums the messages, normalizes the aggregate, applies Lorentz batch normalization, and adds a residual connection; ReLU and a further normalization produce the next state.

Building-block descriptors follow the atom-to-unit assignment. Each atom receives a two-dimensional metal-containing/organic unit indicator and the mean of the 92-dimensional CGCNN descriptors within its unit. These descriptors are broadcast to constituent atoms, so message passing remains at the atomic level. The Lorentz input concatenates learned projections of the atom/type features and the block mean before norm clipping and exponential mapping.

Both branches use the periodic atomic-neighbor graph in the evaluated implementation. The Euclidean branch uses geometric edge embeddings; the Lorentz branch uses BB-aware node inputs and Lorentz transformations. We will add these explicit message/update definitions to Section 4 and clarify the feature construction and graph inputs in Figure 1.

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

This paper introduced GeoDuet, a graph representation model that couples Euclidean and hyperbolic geometry. A Euclidean branch encodes periodic local interactions, while a Lorentz branch processes an atomic covalent graph enriched with buildingblock features. Graph-level attention combines the two representations for property prediction. Building blocks augment atom features rather than define a separate coarse-grained graph.

Critical strengths

The problem setting is valuable: The gas adsorption prediction of MOF materials is of great significance for the screening of porous materials, and the task has practical application value.

Critical weaknesses

The relationship between hyperbolic geometric motivation and performance is not strong enough (Table 9), which makes the "Hierarchical "claim of "Hierarchical Representations" in the title lack strong empirical support. It is suggested that the author make a more explicit distinction between "geometric motivation" and "performance attribution".

Questions for authors

Please compare with the existing Foundation model based MOF gas adsorption and accessible surface area prediction methods.

1 comment

Thank you for the suggestions on the geometric motivation and foundation-model comparisons. We clarify the hierarchy represented by GeoDuet and provide new results for adsorption prediction and surface-area transfer.

1. Hierarchical representation and predictive performance.

The hierarchy in GeoDuet is the atom-to-building-block organization of MOFs. Building-block membership and descriptors are mapped to constituent atoms, allowing the Lorentz branch to incorporate chemical-unit context alongside the Euclidean branch's periodic local geometry. Message passing remains at the atomic level, and the two graph-level representations are combined for prediction. We will make this meaning of hierarchy explicit in the abstract and method description.

Table 9 asks whether graph hyperbolicity predicts per-structure gains, with Spearman correlations of 0.015 for CH4 and 0.082 for CO2. We distinguish this structural diagnostic from the predictive contribution assessed through component comparisons.

In the added experiments, GeoDuet achieves lower MAE than the capacity-matched dual-Euclidean alternative on both tasks: 0.3839 versus 0.4050 for CH4, and 0.7403 versus 0.7682 for CO2. The replacement retains the graph inputs, building-block features, and fusion mechanism, with fewer than 0.2% difference in parameter count. Both models have same layers and use the same 20,000/13,765/13,765 train/validation/test structures and seed 17. The Euclidean and GeoDuet use the fixed 50-epoch schedule. These are comparisons of complete model configurations.

The evaluated branches share the periodic atomic-neighbor graph. The Euclidean branch uses geometric edge features, and the Lorentz branch uses building-block-aware node features with Lorentz transformations. We will clarify these graph inputs in Section 4 and Figure 1 and distinguish the covalent-graph diagnostic from the model's message-passing graph. The component results support the practical value of the coupled representation for prediction.

2. Foundation-model comparison for adsorption and surface area.

We added the official pretrained MOFTransformer, using its crystal-graph and energy-grid inputs. GeoDuet's frozen representations achieve lower MAE on all three targets in the shared-ridge comparison below.

TargetMOFTransformer + ridgeGeoDuet full-data encoder + ridge
CH4, 4.5 bar (mol/kg)0.74520.4269
CO2, 2.5 bar (mol/kg)1.51100.9790
MofDB surface area (m²/g)472.07383.14

Both encoders are frozen. Ridge regression uses alpha = 1, training-set feature standardization, and the same 20,000/13,765/13,765 train/validation/test structures. One structure unsupported by MOFTransformer's preprocessing was excluded from both training sets. GeoDuet's encoders were trained on 110,121 adsorption-labelled structures; MOFTransformer uses its released pretrained weights. The comparison shares the readout and downstream split, while upstream supervision differs. The surface-area row uses the CO2-trained GeoDuet encoder.

We evaluated transfer from the frozen full-data CO2 encoder to surface-area prediction with a fresh attention-based readout. Using 20,000 surface-area training labels, it achieves MAE 153.23 m²/g and R² = 0.9786 on 13,765 test structures, with a separate 13,765-structure validation set and seed 17. Surface area here denotes the MofDB surface_area_m2g field; this supplementary transfer task is reported separately from the accessible-surface-area benchmark in Table 10.

These results demonstrate useful transfer from an adsorption-trained representation to surface-area prediction, alongside lower adsorption errors in the shared-ridge comparison. We will include the pretrained-model protocol and the transfer results in the revision.

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

The authors introduce GeoDuet for predicting gas adsorption and accessible surface area of MOFs (Metal–Organic Frameworks). The model uses two GNN branches; 1) one process local atomic geometry in Euclidean space, 2) the other processes covalent connectivity in Lorentz space. Because secondary building units (organic linker, metal node) of MOFs repetedly exist in a unit cell of MOFs, feature of building blocks is augmented to atomic inputs. They report improvements over the baselines and analyze whether the graph geometry of MOFs is related to these improvements.

Critical strengths

1. combination of spatial and chemical information in encoding

The model GeoDuet combines a Euclidean encoder for local spatial neighborhoods with a Lorentz encoder for covalent connectivity. The model with the combined encoder outperforms the evaluated baselines on adsorption and accessible surface area prediction of MOFs. The authors also analyze the hyperbolicity of MOF covalent graphs, although the correlation between hyperbolicity and performance improvement is not clear.

Critical weaknesses

1. Insufficient evidence or ablation for the benefit of hyperbolic geometry

  1. dual Euclidean baseline with matched graph inputs and comparable capacity is missing.
  2. graph hyperbolicity has little association with prediction gains.
  3. baselines are outdated and relative recent strong baselines (e.g., UniMOF) are in appendix. In addition, baseline training and evaluation protocols are not provided.
  4. comparison of the GeoDuet architecture with more recent models that address MOF structures like AtomMOF[1], Mofasa[2] would be helpful if the authors pointed out their architectural novelty or soundness.

2. Insufficient methodological and experimental detail

Key components including the construction of building block features and the Lorentz message-passing transformations are not sufficiently specified. In particular, baseline training methods (details of the from-scratch of MOF specific models) are not clear. GeoDuet reports two different MAEs of 0.164 and 0.223 for CH4 at 4.5 bar in Tables 2 and 12 without clear explanation, making the results difficult to assess.

[1] AtomMOF: All-Atom Flow Matching for MOF-Adsorbate Structure Prediction. NeurIPS. (2026)

[2] Mofasa: A Step Change in Metal-Oragnic Framework Generation. arXiv. 2512.01756. (2025)

Questions for authors
  1. With the same graph inputs and building block features, does replacing the Lorentz branch with a Euclidean GNN of comparable capacity retain the reported gains?

  2. I recommend reporting the baseline training protocols, including additional baselines with strong performance on MOF gas adsorption prediction, such as MOFTransformer [1], RetNet [2], and RetNetXt [3] and evaluating a broader range of pressures with a detailed interpretation of the results across pressure regimes. In particular, please include results obtained using the official pre-trained checkpoints of these baselines if available.

[1] A multi-modal pre-training transformer for universal transfer learning in metal–organic frameworks. Nature Machine Intelligence. (2023)

[2] Gas adsorption meets deep learning: voxelizing the potential energy surface of metal-organic frameworks. Scientific Reports. (2024)

[3] RetNeXt: A Pretrained Model for Transfer Learning Across the MOF Adsorption Space. Journal of Chemical Information and Modeling. (2026)

3 comments

Thank you for the suggestions. We added component comparisons for CH4 and CO2 and an evaluation of the official pretrained MOFTransformer. We clarify the experimental protocols and model construction below.

1. Does a capacity-matched Euclidean replacement retain the gains?

GeoDuet achieves lower test MAE than the dual-Euclidean alternative on both tasks in the added experiments. The alternative replaces the Lorentz branch with a Euclidean gated GNN while retaining the graph inputs, building-block features, first branch, and fusion mechanism. For these two-layer, width-128 models, GeoDuet and the dual-Euclidean variant have 5,578,849 and 5,585,571 trainable parameters, respectively, a difference below 0.2%.

All models below were trained from scratch on the same 20,000 training structures, with 13,765 validation structures and 13,765 test structures, using seed 17. The component references share a 50-epoch training schedule. Values are test MAE in mol/kg; BB denotes building-block features.

ModelCH4, 4.5 barCO2, 2.5 bar
Euclidean only0.55661.1692
Euclidean + BB0.39330.7530
Dual Euclidean + BB0.40500.7682
GeoDuet without BB0.56281.1899
GeoDuet0.38390.7403

GeoDuet has the lowest MAE among these variants on both gases. The Euclidean-only and Euclidean + BB rows also show the benefit of incorporating building-block information under the same reference training schedule. Together, these results extend the component evaluation beyond CH4 and support the predictive value of the complete GeoDuet design.

2. Strong pretrained baselines and pressure coverage.

We added the official pretrained MOFTransformer, including its crystal-graph and energy-grid inputs. With the same ridge readout and downstream data split, GeoDuet's frozen representations yield lower MAE on both adsorption targets.

Model or frozen representationPrediction moduleCH4, 4.5 barCO2, 2.5 bar
MOFTransformer, official pretrained encoderRidge0.74521.5110
GeoDuet, full-data adsorption encoderRidge0.42690.9790
GeoDuet, complete two-layer modelAttention fusion + regression0.38390.7403

The first two rows use frozen encoders and ridge regression with alpha = 1 and training-set feature standardization. The readout train/validation/test sizes are 20,000/13,765/13,765; one structure unsupported by MOFTransformer's preprocessing was excluded from both training sets. GeoDuet's encoder was trained on 110,121 adsorption-labelled structures, while MOFTransformer uses its released pretrained weights.

The main benchmarks cover five CH4 pressures from 0.05 to 4.5 bar, five CO2 pressures from 0.01 to 2.5 bar, and three Ar pressures from 0.001 to 0.1 bar. The new component experiments cover CH4 at 4.5 bar and CO2 at 2.5 bar. We discuss the larger absolute gains at higher pressures in Response 5 below. Ar results use the log(1 + y) scale.

3. Full-data benchmarks and subset-based ablations.

The 0.223 CH4 MAE in Table 12 comes from ablation experiments on a sampled dataset subset; 0.164 in Table 2 is the full-data benchmark result. We will make this distinction explicit and specify Table 12's subset selection and data split. In the full-data setting, we reproduce MAEs of 0.16414 for CH₄ and 0.33001 for CO₂ on 13,765 test structures, consistent with the results reported in Table 2.

The full-data GeoDuet protocol uses 110,121/13,765/13,765 train/validation/test structures, four layers, width 256, 50 epochs, RAdam, batch size 100, learning rate 0.001, zero weight decay, and seed 123. The original Euclidean references use AdamW and batch size 200. Uni-MOF in the appendix is trained from scratch; MOFormer uses a separate test split. We will provide per-model protocol details alongside the results and distinguish them from the supplementary settings in Q1 and Q2.

4. Building-block features and Lorentz message passing.

Each atom has a 92-dimensional CGCNN descriptor. Its building-block features comprise a two-dimensional metal-containing/organic unit indicator and the mean atomic descriptor within its assigned unit. Learned projections of the atom/type features and the block mean are concatenated, norm-clipped, and mapped to Lorentz space. This incorporates building-block context into the atomic representations.

For each neighboring pair, Lorentz-concatenated features enter separate gate and value transformations, each comprising two Lorentz fully connected layers with intermediate batch normalization and SiLU activation. The channel-wise sigmoid gate weights the value. Messages are summed over neighbors and normalized, followed by a batch-normalized residual update, ReLU, and normalization.

Both branches use the periodic atomic-neighbor graph in the evaluated implementation. The Euclidean branch encodes distances and edge–lattice orientations, while the Lorentz branch uses building-block-enriched node features and Lorentz transformations. We will make the descriptor construction and message/update functions explicit in Section 4 and clarify the graph inputs in Figure 1.

5. Pressure-dependent gains and architectural positioning.

The full-data benchmarks show larger absolute MAE gains at higher pressures. Against the best reported baseline at each pressure, CH4 improves by 0.002 mol/kg at 0.05 bar versus approximately 0.016 at 4.5 bar; CO2 improves by 0.004 at 0.01 bar versus approximately 0.015 at 2.5 bar, using the reproduced CO2 result. This is an absolute-error trend; percentage reductions are not monotonic. Table 9 examines graph hyperbolicity versus per-structure gains, a separate question from this pressure trend.

GeoDuet couples periodic Euclidean geometry with building-block-aware Lorentz representations through graph-level attention for property regression. AtomMOF predicts joint MOF–adsorbate structures, while Mofasa generates frameworks. GeoDuet addresses adsorption and property prediction through complementary atomic representations, with its empirical value supported by the benchmark and component comparisons above.

Thank you for the new experimental results. However, most of my concerns remain unresolved.

In the table in your first response, GeoDuet improves over Euclidean + BB by only 2.39% for CH4 and 1.69% for CO2. It is too marginal. Although you didn't report any seed variance or confidence interval, I suspect that differences of this magnitude could fall within the variance level.

The baseline coverage is still far too limited. As the author-reviewer discussion period will end soon, I won't ask for additional baseline comparisons, but there exist many MOF adsorption uptake prediction models (e.g., RetNeXT, RetNet, DeepSorption, MADField, ...). Considering the time limit, I suggested adding RetNet and RetNeXT and using a pre-trained MOFTransformer encoder. But the authors only added results for MOFTransformer, which was published in 2023. RetNeXT and RetNet are super-light models. These have only around 1M parameters.

I therefore ran a small independent experiment with RetNeXT on just one condition, CH4 at 4.5 bar:

ModelTest MAE (mol/kg) ↓
GeoDuet (reported in your response)0.16414
RetNeXT (my independent experiment)0.16073

I used the same train/validation/test counts, but could not verify that the structure IDs matched your split. This is a single-seed result, so I do not claim a definitive performance advantage, but it reinforces my concern about the omission of this baseline.

The points below are minor issues.

I do not understand the motivation for predicting surface area with an ML model. It can be calculated directly using software like Zeo++, or a GPU implementation of the algorithm is easy and fast. Calculation of surface area is computationally really cheap compared to GCMC simulation (for gas uptake).

The hMOF database includes CH4 adsorption data at 35 bar. For CH4 storage (and in most MOF application cases), adsorption at higher pressure is particularly important. I wonder why authors exclude this condition in training set.

For Ar adsorption, the authors state that the targets were log-transformed, but the reason is not stated. The Ar uptake data are provided in cm3/gcm^3/g. Simply converting it to mol/kg will make the scale similar to other uptake data. Similarly, the CH4 data at 35 bar are stored in cm3/gcm^3/g, whereas the other pressure conditions use mol/kg. This raises the question of whether the 35 bar condition was excluded because of the difference in units.

There are many such small but consequential omissions throughout the paper, and the main claims remain insufficiently supported. I therefore maintain my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.