acceptodds
Withdrawn from ICLR 2027

ArchEGraph: A Large-Scale Graph Dataset for Geometry–Topology–Physics Aligned Building Energy Modeling

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Learning transferable surrogates for building performance requires data that expose how geometry and spatial topology interact with environmental forcing, yet existing datasets rarely align these modalities at scale. We present ArchEGraph, a large-scale graph dataset aligning building geometry, topology, weather, and zone-level physical responses across 5,481 buildings and 49,326 validated building–weather simulations. The target is the EnergyPlus-derived external thermal load under fixed constructions and internal schedules, isolating variation attributable to geometry, topology, and weather rather than operational energy use. ArchEGraph supports two structured learning problems: (i) Mesh-to-Graph reconstruction, which infers local adjacency and globally consistent latent space partitions from unordered geometric primitives; and (ii) Graph-to-Energy prediction, which maps a static heterogeneous graph and an 8,760-step weather sequence to node-aligned response trajectories. Standardized protocols evaluate in-distribution performance and generalization across building and climate shifts. Results expose substantial open challenges: current set encoders recover local face adjacency reliably but struggle with global face–space incidence, particularly on irregular manually designed buildings. Moreover, shuffling only face–space assignments within the same architecture increases prediction error, directly demonstrating that correct topology provides useful information beyond model capacity. ArchEGraph thus provides a controlled benchmark for studying structured physical learning and transferable geometric inductive biases.

Primary area:
datasets and benchmarks
Keywords:
Geometry-Topology-Physics Alignment, Heterogeneous Graph Learning, Building Energy Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_ybRMOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces ArchEGraph, a dataset aligning building geometry, spatial topology, weather data, and zone-level hourly external thermal loads. It contains 5,481 building models and 49,326 EnergyPlus simulation cases generated under fixed thermal and operational settings to study how geometry and weather affect thermal loads. The authors define two benchmark tasks: reconstructing building topology from geometry (M2G) and predicting thermal loads from building graphs and weather sequences (G2E). They provide baselines and evaluation protocols covering task performance, generalization across buildings and climates, and a topology ablation that shuffles face–space assignments.

Critical strengths
  1. The paper is clearly written, with well-designed figures that explain the data construction pipeline, graph representation, and task definitions. This makes the methodology easy to follow and assess.
  2. The dataset aligns building geometry, spatial topology, weather, and zone-level hourly thermal loads, with benchmark tasks and evaluation splits. This provides a reusable resource for studying geometry- and weather-driven thermal loads and comparing methods under consistent conditions.
Critical weaknesses
  1. The dataset’s distinct value and practical significance are not sufficiently established. Its contribution mainly involves integrating and scaling existing geometry processing, graph representation, and simulation workflows, with envelope constructions, HVAC settings, internal loads, and schedules held fixed. The two benchmarks do not adequately demonstrate what new research or applications this controlled resource enables beyond existing modeling and simulation workflows, leaving the significance of the dataset contribution insufficiently justified.
  2. The advantage of graph modeling over conventional methods is not sufficiently demonstrated. WeatherMLP uses only weather inputs, so comparisons against it conflate the benefits of additional geometric information with those of graph structure. The topology-shuffling ablation shows that correct face–space assignments help the current graph model, but the reported evaluation still lacks non-graph baselines using both geometry and weather, such as gradient-boosted trees or an MLP incorporating zone volume, surface area, orientation, and window-to-wall ratio. Without these comparisons, the additional benefit of graph modeling remains unclear.
Questions for authors
  1. Please specify the dataset’s intended use case and demonstrate its benefit over existing automated modeling and EnergyPlus simulation workflows through a concrete application. If the main benefit is speed, report simulation, training, and inference costs, and clarify the scope of these benefits under the fixed thermal and operational settings.
  2. Please report non-graph baselines that use both geometry and weather. The public repository already lists GeometryHistGBR, DegreeHourRidge, and RCInspiredRidge, but their results are not reported in the paper. Compare them with graph models using comparable inputs and identical splits, including cross-building and cross-climate tests, to establish the additional benefit of graph modeling.
  3. Please clarify the IDF generation pipeline in Appendix A.2.4: are IDF files generated directly from the released graph files, or do they depend on an internal building model that retains complete geometry? Provide a reproducible example showing how geometric fields, space assignments, and semantic attributes map to IDF objects. List any additional dependencies, such as raw meshes and fixed templates, to substantiate the claimed loss-preserving mapping.
  4. Please clarify whether G2E uses graphs obtained through preprocessing or graphs predicted by M2G. If the former, report downstream results using M2G predictions. In Table 3, the best face–space incidence F1 scores are approximately 0.43 and 0.25 on the P and M subsets, respectively. Quantifying how these reconstruction errors affect load prediction is necessary to assess the two tasks as a combined pipeline.
  5. Please explain how graph nodes are matched to thermal-load labels. Appendix A.2.5 describes truncation or zero-padding when dimensions differ, but matching tensor shapes does not ensure correct zone correspondence. Report the number of affected samples, the matching rules, and whether padded entries are masked during training and evaluation.
  6. Please provide the exact procedure for computing external thermal disturbance load labels, including the EnergyPlus output variables, sign conventions, temporal aggregation, and area normalization. The load decomposition formulas in the appendix are insufficient to reproduce the labels or verify that they represent the physical quantity defined in the paper.
  7. Please check the units, inverse normalization, and aggregation of the metrics in Table 4. For example, WeatherMLP has an MAE of 8.845 and an RMSE of 6.694 on the P subset. Under the appendix’s definitions, MAE should not exceed RMSE when both use the same units and weights. Please explain this inconsistency and correct the relevant results.
  8. Please clarify the split used for cross-climate evaluation. Section 5.3 states that training and testing use disjoint ASHRAE climate zones, whereas the public repository specifies disjoint weather location IDs and explicitly notes that climate zones may overlap. Identify the protocol used for Table B.6, list the locations and climate zones in each partition, and reconcile the descriptions in the paper and code.
Reviewer_vAjVOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents ArchEGraph, a benchmark dataset for geometry–topology–physics coupled learning in building energy modeling. Buildings are represented as heterogeneous graphs (face nodes, space nodes, face–face and face–space edges), and each building is paired with up to 64 hourly weather sequences from 64 cities, with labels generated by EnergyPlus. The dataset reportedly contains 5,481 buildings (580 manually designed, 4,901 parametric) and 49,326 building–weather cases. Two tasks are defined: M2G (reconstruct building topology from polygon meshes) and G2E (predict zone-level hourly load from graph + weather). Baselines include MLP/DeepSets/Perceiver/SetTransformer/TopoTransformer for M2G and WeatherMLP plus F2S variants for G2E, with cross-building and cross-climate splits. Reported findings are that graph-aware models outperform the weather-only baseline and that shuffling face–space incidence degrades G2E.

Critical strengths

The dataset appears to address a real gap: existing resources seem to have either 3D geometry without physical response labels or energy time series without geometry/topology. Aligning mesh geometry, a space–surface graph, weather forcing, and zone-level simulated loads in one schema could be a useful artifact for structure-aware surrogate modeling.

The benchmark is reasonably well scoped methodologically, with two defined tasks, standardized baselines, and a topology-shuffling ablation that appears to provide a simple check that face–space incidence carries predictive signal.

Critical weaknesses

The scale may be somewhat inflated, and the split protocol is unclear with respect to leakage. The 49,326 cases come from re-pairing a smaller number of distinct buildings with weather files, and it is not clear from the main text whether the 70/15/15 split is over cases or buildings. If it is over cases, the same geometry could appear in both train and test. Building-disjoint splits appear only in the appendix, so it would help to clarify which main-table numbers are leakage-free.

The G2E target and the M2G task may be weaker than the framing suggests. G2E predicts an external disturbance load rather than end-use energy under a single fixed construction template, so the extent to which the task reflects realistic building physics is unclear. It is also not clear how much of the gain comes from topology versus simple envelope statistics. For M2G, the graph construction appears to follow a deterministic rule-based pipeline, and it is not obvious how much genuinely new learning the task requires, nor whether M2G quality translates into G2E accuracy.

Questions for authors
  1. Could you clarify whether the main M2G/G2E tables use case-level random splits or building-disjoint splits, and whether near-duplicate parametric variants are separated across splits?

  2. How much of the G2E variance is explained by weather plus simple envelope statistics without any graph structure? This would help quantify the marginal value of topology.

  3. Does better M2G reconstruction correspond to lower G2E error?

  4. Why is the target the external disturbance load rather than total zone load or HVAC energy? Could total load also be released?

Reviewer_61tCOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

In this paper, the authors focus on the problem of building energy modeling (BEM). They argue that BEM is a graph-based prediction problem for which needs to integrate building elements (zones/units), geometric relations among them, and weather information. However, there is a lack of large-scale dataset that includes all required information for effectively analyzing the performances of graph-based BEM models. The paper evaluates two prediction tasks on the introduced dataset: (1) mesh to graph construction and (2) graph to BEM prediction.

Critical strengths
  • ML for BEM is an important problem.
  • The introduction of a large-scale dataset is valuable.
  • The inclusion of 64 cities is definitely helpful as different cities have different weather profiles.
Critical weaknesses
  1. Not sure how common is the need to estimate zones from a mesh representation. Are zones generally not annotated already? If the authors are using a rule-based method for forming the ground truth for the mesh-2-graph task, it is not clear how/why this task of the dataset is helpful for the community.

  2. The paper includes experimental analysis with a limited number of papers. Including more methods (especially recent ones) would be helpful for forming a helpful baseline.

  3. The paper is missing citations to some relevant literature. Some of these papers have codes, which would contribute to the analysis as well.

Halaçlı, E. G., Canlı, İ., İşeri, O. K., Yavuz, F., Akgül, Ç. M., Kalkan, S., & Dino, İ. G. (2023, November). A novel graph neural network for zone-level urban-scale building energy use estimation. In Proceedings of the 10th ACM International Conference on Systems for Energy-Efficient Buildings, Cities, and Transportation (pp. 169-176).

Reveshti, A. M., Dabbagh, J., Reveshti, J. N., Bonab, K. F., & Mansoub, F. H. (2026). Graph Neural Network‐Based Prediction of Building Energy Consumption. Energy Science & Engineering.

  1. Table 4 includes MAE values that are larger than RMSE errors, which should not be possible.

  2. Are the training-validation-test sets based on non-overlapping buildings and weather profiles? If yes, this should be stated explicitly as a severe limitation. Such a benchmark should be providing splits that are based on non-overlapping buildings and weather profiles as well.

  3. Not sure why the paper applies MLP only on the weather data. For MLP to function as a suitable non-graph-based approach, it should be provided physical properties of the zone as well (without the information about the neighboring zones). As it is, the paper's main motivation for advocating a graph-based dataset is undermined.

  4. It is not clear how many seeds are used.

Questions for authors

Authors are encouraged to comment about the issues listed in the weaknesses part above.

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

Related papers

Discussion (0)

Sign in to comment.