All projects

Machine Learning

CNN-Based Daylight Autonomy Prediction

Real-time daylight autonomy at ~3% MAE, ~20x faster than Radiance

A U-Net CNN that predicts daylight autonomy maps in real time. I owned the 100,000-case synthetic dataset, the Azure ML inference stack, and the Grasshopper interface, used on 30+ projects.

WHY A SURROGATE

Early-stage daylight consultancy is the right moment to influence facade and massing, but Radiance raytracing is too slow for the iteration loop. Teams skip the analysis or run two or three scenarios where they should run hundreds. The architectural cost shows up later as facade rework and shading retrofits.

DATASET

Parametric Ladybug Tools + Radiance pipeline, then 100,000+ synthetic daylight autonomy cases. Sister dataset to the overheating GNN, same parametric Grasshopper base, independent parameter sweep:

  • UK climate files (4 UK climate zones)
  • Single-room rectangular geometries (3-10m depth, 3-8m width, 2.4-3.6m height)
  • Glazing ratios 10-80%
  • External shading (overhangs, fins, none)

Targets are spatial Daylight Autonomy maps at 0.5m grid resolution.

MODEL

U-Net convolutional architecture, co-designed with one colleague. Image-like room encodings in (multi-channel tensors for geometry, glazing, shading, sky), daylight autonomy maps out.

Feature engineering took several iterations: encoding rooms-with-glazing-and-shading into channels the network could learn from was the main blocker.

VALIDATION

  • 3% MAE against Radiance (rpict, ambient_bounces=6) on a held-out test set.
  • ~13 seconds inference vs ~8 minutes per Radiance run on the same room (an Azure ML GPU instance for both, comparable run conditions).
  • Generalises to all four UK climate zones without retraining.

DEPLOYMENT

Azure ML endpoint with GPU backing, FastAPI service, Docker, Grasshopper UI delivering results in seconds inside the architect’s working environment. Request batching to hit the latency target.

ACHIEVEMENTS

  • 3% MAE vs Radiance ground truth, ~13 second inference.
  • 30+ projects using the daylight tool in production design workflows.
  • End-to-end ownership: dataset, inference, deployment, UI.
  • Pattern reused by the later peak-solar and overheating tools.