A multi-modal dataset for learning-based morphology-informed building performance modeling

Area of Science:

• Computational Engineering
• Building Science
• Data Science

Background:

• Building performance is dictated by morphology, but high-fidelity simulations are computationally intensive.
• Existing datasets lack geometric detail and validated simulation outputs, hindering data-driven design.
• Current research relies on limited, non-public datasets, restricting scalability and integration.

Purpose of the Study:

• Introduce PRISM, a public dataset linking diverse building geometry with high-fidelity simulation outputs.
• Enable scalable, data-driven research in building performance and design.
• Facilitate integration of performance-informed design into computational workflows.

Main Methods:

• Created a dataset with high morphological diversity, from simple extrusions to complex forms.
• Included multi-modal geometric representations: watertight meshes, signed distance fields, and structured descriptors.
• Generated cross-domain simulation outputs: CFD (velocity, pressure) and solar exposure (sky view factor).

Main Results:

• PRISM offers a systematic link between building shape and validated physical simulation data.
• The dataset encompasses diverse geometries and multi-modal representations.
• Includes comprehensive environmental and physical simulation outputs.

Conclusions:

• PRISM addresses the need for high-quality, accessible datasets in building performance research.
• It supports various applications like surrogate modeling, inverse design, and generative modeling.
• The dataset promotes large-scale design-space exploration and data-driven design workflows.