PaperlayerView exact revision on arXiv

Source-linked AI summary

A SegFormer-based model for spatial Daylight Factor prediction with comprehensive obstruction encoding and regulatory validation

Stanislava Fedorova, Alejandro Pacheco Diéguez, Christoph Berkmiller, Alberto Floris, Libny Pacheco, Marie-Claude Dubois

arXiv:2610.33737v1cs.CVcs.LG

TL;DR

Daylight compliance assessment requires time-consuming simulations and specialist model preparation, slowing design iteration. This paper develops BIM-native machine-learning models for spatial Daylight Factor prediction and evaluates their accuracy against Radiance and real-building cases. The default-materials model shows strong real-world compliance-zone accuracy, while the authors recommend it for preliminary assessments and identify further development needs for the custom-materials variant.

  • Problem

    Daylight simulations and specialist BIM model preparation take minutes to hours and several hours, respectively, slowing validation of design changes.

    Source: §1 · p. 2

  • Method

    A modified SegFormer predicts spatial Daylight Factor maps from BIM-derived inputs, with separate default-materials and custom-materials variants.

    Source: §3.2 · p. 5

  • Results

    On real buildings, the default-materials model achieves Compliance IoU of 0.883, and all overestimated-compliance cases are borderline to the BBR or TEK17 threshold.

    Source: §6 · p. 12

  • Takeaways & Limitations

    The default-materials model is recommended for preliminary daylight simulations, while the custom-materials model should not replace final compliance simulation.

    Source: §6 · p. 12

    The encoding cannot capture curved windows, tilted walls, skylights, or organic forms, which require traditional raytracing.

    Source: §5.4 · p. 12

Abstract

from arXiv · show

Daylight compliance assessment in residential buildings requires time-consuming raytracing simulations and specialist model preparation, limiting its integration into iterative architectural design. We present a machine learning-based methodology for predicting the spatial distribution of the Daylight Factor (DF) directly within Building Information Modelling (BIM) environments, delivering results in seconds without model export or manual curation. The approach addresses five gaps not jointly covered by any existing model: external obstruction encoding including balconies, own building height, variable material reflectances, non-orthogonal room geometries, and regulatory validation. Two models are developed: a default-materials model with fixed reflectances for preliminary design checks, and a custom-materials model with variable project-specific reflectances. Both are trained on about 20,000 synthetic Radiance-simulated rooms and validated on a held-out synthetic dataset and 102 rooms from 18 real building models. The default-materials model achieves $\delta_1$ = 83.4% and MAE = 0.11% DF on synthetic data, and on real buildings predicts BBR (Swedish) compliance area within 4.85 pp MAE (bias +2.31 pp) and TEK17 (Norwegian) mean room DF within 0.304% DF MAE (bias +0.287% DF). The custom-materials variant shows 1.7-2.4$\times$ higher error across all metrics. Both models concentrate accuracy in the BBR compliance-critical zone (DF < 1%), achieving a sub-1% DF MAE of 0.105% DF and Compliance IoU of 0.883 on real buildings. All cases where compliance is overestimated are borderline, within 10.7 pp of the BBR threshold (median 3.2 pp) and 0.36% DF of the TEK17 threshold (median 0.10% DF). The default-materials model is recommended as a replacement for preliminary daylight simulations; the custom-materials variant requires further development before use in final compliance verification.

1 Introduction

Daylight compliance checks rely on staged simulations, but BIM model preparation and simulation time slow design decisions. This study develops and validates a machine-learning model that predicts Daylight Factor directly in BIM for potential compliance checks.

1 Introduction

  • Swedish BBR requires DF of at least 1.0% over at least 50% of assessed habitable-room floor area, while Norwegian TEK17 requires mean room DF of at least 2.0%.Both regulations also accept simplified assessment methods within defined scopes, with documented accuracy limitations.Source: §1 · p. 2
  • VSC is used early to assess incident window light, but cannot guarantee DF compliance because it omits room, window, glazing, and reflectance factors.The workflow then uses preliminary simulations with default materials and detailed simulations with project-specific materials.Source: §1 · p. 2
  • Model preparation can take several hours of specialist work, while Radiance simulations take minutes to hours, slowing validation of design changes.Faster computation tools do not remove the preparation bottleneck.Source: §1 · p. 2
  • The study develops and validates BIM-native machine-learning models to predict Daylight Factor without model export.The models are intended to assess whether their accuracy supports preliminary or final compliance verification.Source: §1 · p. 2

2 State of the Art in ML-based Daylight Prediction

Prior daylight-prediction models leave gaps in obstruction, building, material, geometry, and regulatory coverage. This study addresses all five together and validates the resulting model against real-world building models.

2 State of the Art in ML-based Daylight Prediction

  • Some models omit surrounding context, while others omit balconies or distant buildings, limiting obstruction coverage in settings where lower-floor rooms may be most affected.Pacheco Dieguez et al. encoded surrounding obstructions using a window-specific decomposition strategy.Source: §2 · p. 3
  • None of the reviewed models encodes host-building height, a factor relevant to daylight in dense urban contexts.Lower floors may be surrounded by obstructions on multiple sides.Source: §2 · p. 3
  • Most reviewed models use fixed reflectances, while Li et al. includes reflectance inputs but omits external obstructions and regulatory validation.Fixed reflectances may suffice for preliminary checks but not final official verification requiring project-specific inputs.Source: §2 · p. 3
  • All reviewed models, including, are restricted to orthogonal rooms and rectangular windows.Source: §2 · p. 3
  • Few studies evaluate performance against compliance thresholds, and aggregate metrics such as MAE or R² do not show whether rooms are classified correctly.The literature spans methods from ANNs and SVMs to CNNs, GANs, and ensembles.Source: §2 · p. 3
  • This study addresses all five identified modelling gaps simultaneously and validates its model against real-world building models.The gaps concern external obstructions, host-building height, variable reflectances, non-orthogonal geometries, and regulatory validation.Source: §2 · p. 3, Table 1 · p. 3

3 Methodology

The methodology trains a SegFormer-based model on Radiance-simulated room maps and evaluates it on held-out synthetic and real-building data. It encodes room geometry, windows, obstructions, and selected project parameters, then assesses spatial accuracy and regulatory compliance.

3.1 Dataset and ground truth generation

  • The 20,687-room synthetic dataset covers rectangular and L-shaped rooms, varied windows and balconies, and urban contexts, with window sizes focused around the Swedish DF ≥ 1% threshold.Two variants use either fixed standard reflectances or variable project-specific reflectances.Source: §3.1 · p. 4
  • Radiance ground truth is generated on a 0.25 × 0.25 m grid at 0.8 m height, and evaluation uses 500 held-out synthetic rooms plus 102 rooms from 18 real building models.The real buildings span five urban typologies and varied geometries, windows, obstruction contexts, and floor heights.Source: §3.1 · p. 4

3.2 Model architecture

  • A modified SegFormer with a Mix Vision Transformer encoder and Global-Local Path decoder produces a 384 × 384 DF map using an L1 loss.The output is limited to the 0–9.9% DF range.Source: §3.2 · p. 5

3.3 Input encoding

  • Each window’s daylight contribution is predicted independently and summed in postprocessing, reducing the learning task to single-window predictions and allowing complex obstructions to be included.Inputs combine a room-and-window geometry mask with numerical conditioning parameters; obstruction angles span 64 planes across a 145° field of view.Source: §3.3 · p. 5
  • When a sector has no visible sky, both obstruction angles are set to 45°, making fully blocked sectors indistinguishable despite differences in balcony-building overlap.The authors note that this distinction can substantially affect daylight levels, although real-world results retain good overall accuracy across obstructed configurations.Source: §3.3 · p. 5

3.4 Validation metrics

  • Validation includes threshold accuracy, quantized IoU, MAE, and regulatory measures for BBR and TEK17, with sub-1% DF error isolated around the BBR boundary.Pixels crossing the 1% DF threshold in either direction count as direct compliance errors.Source: §3.4 · pp. 5–6

4 Results

The default-materials model performs well on synthetic and real-building validation, including the sub-1% DF zone used for BBR compliance. On real buildings, its BBR area and TEK17 mean-DF estimates show slight positive bias, with overestimation confined to borderline cases.

4.1 Synthetic validation (Dataset 1, n=499)

  • 83.4% of default-material predictions fell within 25% of ground truth (δ₁), and 99.3% within ~95% (δ₃); custom materials performed worse across all metrics.The strongest degradation was in the sub-1% DF zone, where δ₁ fell from 83% to 43% and MAE increased 2.4×.Source: Table 2 · p. 6, §4.1 · p. 6
  • Fig. 2 traces a synthetic room from its geometry-mask input through Radiance ground truth, prediction, signed error, and BBR compliance overlay.The error panel uses a ±2% DF scale.Source: Figure 2 · p. 7

4.2 Real-world validation (Dataset 2, n=102)

  • 0.105% DF sub-1% MAE and 0.883 Compliance IoU were achieved on real buildings, with sub-1% metrics reported for 90 rooms.The full real-world dataset comprised 102 rooms from 18 building models.Source: Table 4 · p. 7, §4.2 · p. 7
  • 4.85 pp MAE (bias +2.31 pp) for BBR compliant-area estimates and 0.304% DF MAE (bias +0.287% DF) for TEK17 mean room DF were reported.Every case of overestimated compliance was borderline: within 10.7 pp of the BBR threshold (median 3.2 pp) and 0.36% DF of the TEK17 threshold (median 0.10% DF).Source: §4.2 · pp. 7–8
  • Real-room examples span deep balconies, buildings in front, combined obstructions, and multi-window layouts.The figure presents five representative rooms and their ground-truth maps, predictions, errors, and BBR compliance overlays.Source: Figure 3 · p. 8, §4.2 · p. 8

5 Discussion

The discussion places the default-materials model’s accuracy in context, examines its regulatory implications and BIM workflow benefits, and identifies boundaries for current use. It recommends further development of the custom-materials model before certification-grade application.

5.1 Accuracy in context

  • On real buildings, the default-materials model’s MAE is 0.348% DF overall and 0.105% DF below 1% DF; the custom model’s sub-1% DF MAE is 0.121% DF.The authors describe the default model as functionally equivalent to traditional simulation for compliance in standard architectural configurations, while the custom model is suitable for exploratory design but not certification-grade reporting.Source: §5.1 · p. 9
  • The five gaps addressed together are external obstructions, host-building height, variable reflectances, non-orthogonal geometry, and regulatory validation against real buildings.The paper presents this combination as absent from previous models.Source: §5.1 · p. 9

5.2 Regulatory compliance and simplified method comparison

  • For the default model, positive compliance bias is concentrated in high-DF zones, leaving BBR area predictions largely unaffected; TEK17 mean DF is biased upward by +0.287% DF.On real buildings, overestimation is concentrated in clearly compliant rooms; false-positive TEK17 cases are within a median 0.10% DF of the 2.0% threshold.Source: §5.2 · pp. 9–10
  • The custom-materials model’s sub-1% DF δ₁ falls to 43%, so it is indicative for material sensitivity but should not replace final compliance simulation.For materials equal to or lighter than the default set, the default model can demonstrate compliance; darker materials require Radiance simulation.Source: §5.2 · p. 10
  • The model’s BBR compliance-area MAE is 4.85 pp and TEK17 mean-room-DF MAE is 0.304% DF, while accepted simplified methods have documented accuracy limitations.The paper argues these results provide a basis for considering the model as a compliance assessment tool.Source: §5.2 · p. 10

5.3 BIM integration and practical benefits

  • BIM-native encoding avoids export, geometry simplification, material assignment, and specialist curation; strong divergence from Radiance in simple rooms can flag import failures.Several such Radiance import failures were identified and confirmed during real-world validation, suggesting a potential independent quality-control check.Source: §5.3 · p. 11
  • Beta deployments with two Scandinavian architecture firms preliminarily estimate an 80–90% reduction in total daylight-compliance time versus traditional workflows.Immediate visual feedback may also help architects develop intuition about how design parameters interact with compliance.Source: §5.3 · p. 11

5.4 Limitations and Future Work

  • Curved windows, tilted walls, skylights, and organic forms fall outside the current encoding and require traditional raytracing; extending beyond static DF is future work.The authors also identify replacing or supplementing synthetic training data with real-building simulations and physical measurements as a way to reduce the domain gap.Source: §5.4 · p. 12

6 Conclusions

The study presents BIM-native spatial DF prediction and reports regulatory validation on real buildings. The default-materials model is recommended for preliminary assessments, while Radiance remains necessary for certification and specified non-standard configurations.

6 Conclusions

  • Five gaps are addressed jointly: external obstructions including balconies, own building height, variable reflectances, non-orthogonal rooms, and regulatory-threshold validation.The paper describes this combination as addressed for the first time.Source: §6 · p. 12
  • BIM-native encoding removes model export, manual curation, and specialist preparation, while seconds-long parallel predictions support whole-building assessments at any design stage.The workflow also enables detection of Radiance simulation errors as an independent quality check.Source: §6 · p. 12
  • On 102 real rooms, the default-materials model achieved a pixel-level median error of 0.289 % DF and Compliance IoU of 0.883; overestimated compliance cases were borderline.They were within median 3.2 pp of the BBR threshold and median 0.10 % DF of the TEK17 threshold.Source: §6 · p. 12
  • The default-materials model is recommended for preliminary daylight simulations, including VSC checks and early-stage DF assessments.It is also recommended for final compliance verification when project materials are equal to or lighter than the default set.Source: §6 · p. 12
  • Radiance simulation is required for certification; non-standard configurations—including large rooms, double-height spaces, curved or organic forms, and tilted windows or walls—also require Radiance until models covering these conditions are developed.Source: §6 · p. 13
  • Immediate design-stage feedback is intended to help identify likely non-compliance early, so final-simulation discrepancies may be minor and localised rather than require fundamental redesign.The paper also argues that rapid verification within the design environment supports the protection of daylight regulation and architects’ efforts to balance density with adequate daylight.Source: §6 · p. 13

Contributions and Acknowledgements

The contribution statement assigns leadership across model development, project coordination, BIM integration, validation, and writing; acknowledgements identify funding and external building-model contributions.

Contributions and Acknowledgements

  • Fedorova led model architecture, training, and ML experimentation; Pacheco Diéguez led problem framing, dataset generation, obstruction encoding, BIM integration, validation, and manuscript writing.They contributed equally and share first authorship.Source: Contributions and Acknowledgements · p. 13
  • Belysningsstiftelsen funded the first default-materials model and early practical testing; Vinnova, ALMI, and Scaleway supported further development and commercial deployment.White Arkitekter and LINK Arkitektur provided real-world building models and beta-testing feedback.Source: Contributions and Acknowledgements · p. 13

AI Disclosure Statement

The authors disclose Claude-assisted manuscript drafting and editing, with no AI use in data analysis, figure generation, or model development.

AI Disclosure Statement

  • Claude Sonnet 4.6 supported manuscript drafting and editing; the authors reviewed and approved the content and take responsibility for it.No AI tools were used for data analysis, figure generation, or model development.Source: AI Disclosure Statement · p. 13

2nd Edition. BRE, Garston (2011).

The appendix describes parameter distributions for the 20,687-room training dataset, covering geometry, windows, envelope and urban context, custom-material reflectances, and simulated DF outputs.

Appendix A — Synthetic dataset: parameter distributions

  • The geometry distributions cover room depth, width, ceiling height, and floor area across 20,687 rooms.The appendix identifies these as input parameters used in both model variants.Source: Figure 4 · p. 16, Appendix A · p. 16
  • Window distributions include width, height, sill height, and area; the sill-height bimodality reflects floor-to-ceiling and standard-window configurations.The stated configurations are sill = 0 m and sill 0.8–1.0 m, respectively.Source: Figure 5 · p. 16
  • Envelope and urban-context distributions include façade thickness, frame ratio, floor height above terrain, and VSC at the window centre.These distributions are shown for the full 20,687-room dataset.Source: Figure 6 · p. 17
  • WWR is shown for all 20,687 rooms and balcony depth for the 9,674 rooms with balconies; the WWR distribution reflects single-window configurations.Effective total WWR is higher in multi-window rooms.Source: Figure 7 · p. 17
  • Custom-material reflectance distributions cover ceiling, walls, floor, façade, surrounding buildings, window frame, terrain, and balcony ceiling.Balcony-ceiling values are for rooms with balconies only; other surfaces cover all 20,687 rooms.Source: Figure 8 · p. 18
  • The simulated-output distributions show spatial average and median DF per room under a Radiance CIE overcast sky; values above 3% are excluded for readability.Mean and median values are indicated on the histograms.Source: Appendix A · p. 16, Figure 9 · p. 18

Appendix B — Custom materials model: reflectance analysis

The appendices examine reflectance-related error differences for the custom-materials model and summarize the real-world validation dataset and its parameters.

Appendix B — Custom materials model: reflectance analysis

  • Table 5 reports Pearson correlations between surface reflectance and per-sample MAE increase for df_custom relative to df_default across 499 synthetic test samples.The appendix frames this analysis as examining which material parameters drive the error increase.Source: Table 5 · p. 19, Appendix B · p. 19

Appendix C — Real-world validation dataset: summary

  • The real-world dataset comprises 102 rooms from 18 buildings across five urban typologies, after excluding five anomalous rooms.Buildings contribute 2–13 rooms each, with a mean of 5.7 rooms.Source: Appendix C · p. 20
  • Table 6 reports surface reflectance values assigned across the 18 real-world validation buildings for the custom-materials variant.Source: Table 6 · p. 20
  • Tables 7 and 8 summarize room-level parameters for 102 rooms and window-level parameters for 170 windows, respectively.Source: Table 7 · p. 20, Table 8 · p. 20
Loading 2610.33737v1…