This project introduced Per-FORM, an implicit neural representation (INR)–based surrogate modeling framework for predictive modeling of aerodynamics and heat transfer in the built environment. The approach learns a shape-conditioned mapping between compact geometric encodings and continuous physical fields, enabling scalable full-field and near-surface inference across geometries of varying scale and topological complexity. By decoupling representation resolution from geometric complexity, the framework supports flexible modeling of coupled geometry–performance relationships. The model achieved competitive accuracy relative to grid-based baselines while extending prediction from single-plane outputs to continuous field inference. Beyond simulation acceleration, the learned latent space enabled performance-informed design exploration, supporting forward and inverse correspondences between form and environmental behavior.

The work was exhibited at The Next Earth: Computation, Crisis, Cosmology at Palazzo Diedo, Berggruen Arts & Culture, as part of the 19th Venice Architecture Biennale (May 2025).