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

Building performance, including airflow and radiative exchanges, is governed by morphology. High-fidelity simulations are accurate but computationally expensive, limiting their use in design workflows. Data-driven approaches provide a scalable alternative, yet their development remains constrained by the absence of high-quality datasets that pair geometrically expressive 3D models with validated physical simulation outputs. Existing datasets are predominantly derived from footprint extrusions, photogrammetric reconstruction, or LoD1-2 city models, and are limited in their suitability for physics-based simulation workflows. They frequently omit critical morphological features, lack watertightness and topological consistency, and do not include associated simulation fields. As a result, performance-informed, shape-driven research often relies on single-use, non-public datasets tailored to narrowly scoped tasks, severely limiting broader adoption, scalability, and integration into data-driven design workflows. This work introduces PRISM (Performance Representation Informed by Shape and Morphology), a publicly available dataset that systematically links building geometry with high-fidelity simulation outputs across multiple environmental domains. PRISM features high morphological diversity, spanning basic extrusions to highly articulated forms; multi-modal geometric representations, including watertight surface meshes, signed distance fields and structured descriptors; and cross-domain environmental and physical simulation outputs from computational fluid dynamics (velocity, pressure) and visibility-derived solar exposure (sky view factor by sky patch). PRISM supports a wide range of research applications, including surrogate modeling, geometry-conditioned performance prediction, inverse design, generative modeling, and large-scale design-space exploration.