Room parameters
Predicted illuminance
How it works
The network was trained on 1,000 Radiance simulations of a room fitted with SerraGlaze film, using the film's measured BSDF, and tested on 100 held-out simulations. Each simulation records illuminance on a 13 × 13 grid of sensors across the room.
The four room parameters are standardised and passed through a small fully connected network (4 → 128 → 256 → 128 → 20, ReLU). The network predicts 20 principal components of the illuminance map, which are projected back to the 169 sensor values. These components capture 99.3% of the variance in the training data. The same weights and preprocessing as the original PyTorch model run here in plain JavaScript, so nothing is sent to a server.
Results
| Model | Predicts | R² | Speed-up over Radiance |
|---|---|---|---|
| Neural network (PyTorch) | full 13 × 13 illuminance grid | 0.988 | ~4 × 10⁵ |
| XGBoost | average room illuminance | 0.986 | ~7 × 10⁶ |
This page runs the neural network. The XGBoost model only predicts the average, so it isn't included here.
Limitations
- Validated against Radiance simulations, not physical measurements.
- Azimuth is not encoded as periodic, so predictions at 0 and 2π differ by about 7% even though they are the same direction.
- At some extreme settings the network predicts slightly negative illuminance for a few sensors. These are shown as 0 lux on the map.
Context
Final-year industrial project with SerraLux, University of Oxford Department of Physics, 2025–26. Awarded the Department of Physics prize for best group project presentation.