This project developed a context-aware predictive framework for urban blocks based on a hybrid neural representation that jointly encodes built form and surrounding environmental context. The approach learns a geometry- and context-conditioned mapping between structured urban morphology and continuous environmental performance fields. By integrating discrete structural encodings with continuous latent fields, the framework captures both fine-grained geometric detail and broader spatial dependencies, enabling full-field prediction across heterogeneous blocks and districts.

The proposed framework is aligned with the level at which design agency is most actionable, the building to urban block scale, while remaining extensible toward neighborhood- and city-scale analysis through bottom-up aggregation and abstraction. Beyond forward prediction, the learned latent space enables compositional and performance-informed generative workflows: urban configurations can be recombined, interpolated, and perturbed in a differentiable manner, supporting performance-informed and design-aware exploration and optimization. This allows environmental characterization at scale, as well as multi-scale risk and opportunity mapping that connects local interventions to district-wide outcomes.