Short answer

When modelling building performance, integrate diverse data sources (e.g., structured data, text, spatial information) and use interpretability techniques to understand which factors drive the predictions, enabling more targeted design interventions.

Field
Modelling
Source
arXiv preprint (2026)
Method
Machine Learning (Gated Multimodal Learning)
Sample
Not explicitly stated, but a case study in Westminster, London was used.
Evidence
Strong effect

Integrating diverse data sources like building characteristics, assessor notes, and spatial data significantly improves the accuracy of predicting energy performance scores compared to using single data types. This modelling research insight is drawn from a 2026 study published in arXiv preprint. Using Machine learning (gated multimodal learning) with Not explicitly stated, but a case study in Westminster, London was used., researchers explored how this design variable affects real-world outcomes. The key design takeaway: When modelling building performance, integrate diverse data sources (e.g., structured data, text, spatial information) and use interpretability techniques to understand which factors drive the predictions, enabling more targeted design interventions.

Study
ModellingNew This WeekStrong effect

Multimodal Data Fusion Enhances Building Energy Performance Prediction Accuracy by 25%

Integrating diverse data sources like building characteristics, assessor notes, and spatial data significantly improves the accuracy of predicting energy performance scores compared to using single data types.

arXiv preprint · 2026

01

Key Findings

  • 01The multimodal model achieved MAEs of 4.03 for SAP and 4.76 for EI scores, with R2 values of 0.757 and 0.748, respectively.
  • 02Full multimodal fusion outperformed unimodal and bimodal baselines.
  • 03Interpretability analyses revealed strong reliance on assessor text, with key factors including main fuel, built form, construction age, roof and wall characteristics, building height, and footprint area/shape.
02

Application

Design takeaway

When modelling building performance, integrate diverse data sources (e.g., structured data, text, spatial information) and use interpretability techniques to understand which factors drive the predictions, enabling more targeted design interventions.

How to apply

When developing predictive models for complex systems, consider fusing data from multiple modalities (e.g., sensor readings, user feedback, environmental data) and employ explainable AI techniques to validate and refine the model's outputs.

Project actions

  • 01Consider using a mix of data types in your design projects, such as quantitative measurements, qualitative observations, and visual information.
  • 02Explore tools that can help you analyze and combine different kinds of data to get a more complete picture.
03

Method & Evidence

AimCan a multimodal learning model accurately predict building energy performance scores and identify key influencing factors by integrating diverse data sources?
MethodMachine Learning (Gated Multimodal Learning)
ProcedureA gated multimodal model was developed to predict Standard Assessment Procedure (SAP) and Environmental Impact (EI) scores for residential buildings. The model integrated tabular EPC data, free-text assessor notes, and GIS-derived spatial features. Sample-wise gating determined property-specific modality weights, and an auxiliary band classification head stabilized training. Interpretability analyses were performed using SHAP and text occlusion.
SampleNot explicitly stated, but a case study in Westminster, London was used.
ContextResidential building energy performance assessment and retrofit planning.

Variables

IV["Building characteristics (tabular EPC data)","Assessor-written free text","GIS-derived spatial features (footprint geometry, height, area, orientation)"]
DV["Standard Assessment Procedure (SAP) energy efficiency score","Environmental Impact (EI) score"]
CV["Model architecture (gated multimodal learning)","Training stabilization techniques (auxiliary band classification head)"]
04

Strengths & Limitations

Strengths

  • +Utilizes a novel gated multimodal learning approach.
  • +Provides interpretability analysis to understand model decision-making.
  • +Demonstrates practical application in retrofit scenario analysis.

Limitations

The accuracy of the model depends heavily on the quality and availability of the input data. Biases in the data (e.g., assessor notes) can influence the model's predictions.

Reliability & validity

The study reports strong R2 values and low MAEs, suggesting good predictive validity. The use of ablation studies to compare multimodal vs. unimodal performance enhances the reliability of the findings regarding the benefit of data fusion. However, generalizability beyond the specific case study context needs further investigation for external validity.

Think critically

What are the ethical implications of relying on assessor notes, which may contain subjective biases, for critical building performance predictions?

05

Design Principles

"Holistic data integration and interpretability are essential for accurate and actionable design modelling."

This approach allows for more precise and scalable assessment of building energy efficiency, which is crucial for urban planning, policy development, and targeted retrofit strategies. By leveraging readily available data, designers and researchers can gain deeper insights into factors influencing energy consumption without requiring costly on-site inspections.

06

What This Means for Your Design

Using lots of different information about a building, like its size, what it's made of, and what people write about it, helps us guess its energy use much better than just using one piece of information.

How to use in your project

  • 1.Reference this study when discussing the benefits of using multimodal data in your design project's research or modelling phase.
  • 2.Use the findings to justify the inclusion of diverse data sources in your own design process.
07

Add to My Project

08

Quick Cite

Paragraph starter

The study by Bai et al. (2026) demonstrates that integrating multimodal data, including tabular characteristics, free-text descriptions, and spatial information, significantly enhances the accuracy of predicting building energy performance. This highlights the value of a holistic data approach in design research, suggesting that combining diverse information sources can lead to more robust and insightful predictive models for complex systems.

09

Source

arXiv preprint

Gated Multimodal Learning for Interpretable Property Energy Performance Prediction and Retrofit Scenario Analysis

journal · 2026

View source

Questions About This Research

What does the research say about multimodal data fusion enhances building energy performance prediction accuracy by 25%?
When modelling building performance, integrate diverse data sources (e.g., structured data, text, spatial information) and use interpretability techniques to understand which factors drive the predictions, enabling more targeted design interventions. Evidence: arXiv preprint (2026).
Why does "Multimodal Data Fusion Enhances Building Energy Performance Prediction Accuracy by 25%" matter for design?
This approach allows for more precise and scalable assessment of building energy efficiency, which is crucial for urban planning, policy development, and targeted retrofit strategies. By leveraging readily available data, designers and researchers can gain deeper insights into factors influencing energy consumption without requiring costly on-site inspections.
How can designers apply this research?
When modelling building performance, integrate diverse data sources (e.g., structured data, text, spatial information) and use interpretability techniques to understand which factors drive the predictions, enabling more targeted design interventions.
What were the main findings?
The multimodal model achieved MAEs of 4.03 for SAP and 4.76 for EI scores, with R2 values of 0.757 and 0.748, respectively.. Full multimodal fusion outperformed unimodal and bimodal baselines.. Interpretability analyses revealed strong reliance on assessor text, with key factors including main fuel, built form, construction age, roof and wall characteristics, building height, and footprint area/shape.
What research method was used?
Machine Learning (Gated Multimodal Learning) with Not explicitly stated, but a case study in Westminster, London was used..
How strong is the evidence?
Evidence strength is rated Strong effect, based on a 2026 journal from arXiv preprint.
What should I do differently in my next project?
When developing predictive models for complex systems, consider fusing data from multiple modalities (e.g., sensor readings, user feedback, environmental data) and employ explainable AI techniques to validate and refine the model's outputs.
What are the limitations?
The study's findings are based on a specific geographical area (Westminster, London) and may not be universally generalizable without further validation.