Energy Decay Curve Prediction
Predicting acoustic room characteristics directly from room geometry
Deep learning framework that predicts a room’s acoustic behavior — energy decay curve, reverberation time, clarity — directly from its geometry and materials, as a data-driven alternative to expensive physical measurement or geometric acoustics simulation.
Approach
A multi-head CNN-LSTM network takes 16 room features (dimensions, wall/floor/ceiling absorption, material types) and predicts three outputs jointly: the full energy decay curve (96,000 samples), reverberation time T20, and clarity index C50.
- CNN pathway — 1D convolutions (filters [32, 64], kernel 3) over the feature vector.
- LSTM pathway — bidirectional LSTM (128 hidden units) for sequential context.
- Fusion — concatenated features through shared FC layers (2048 → ReLU → Dropout 0.3) into three output heads.
Trained on 6,000 simulated room configurations (Zenodo dataset) with a weighted multi-task loss (EDC×1.0, T20×100, C50×50) — the heavy weighting on the two scalar heads compensates for the 96,000-dimensional EDC head otherwise dominating the loss. Compared against pure-LSTM and Transformer baselines before settling on this multi-head architecture.
Results
| Acoustic Parameter | MAE | RMSE | R² |
|---|---|---|---|
| EDC (normalized) | 0.000257 | 0.00213 | 0.9995 |
| T20 (seconds) | 0.0647 | 0.1106 | 0.9530 |
| C50 (decibels) | 0.338 | 0.610 | 0.9917 |
103M-parameter model, 200 epochs, 94.2 minutes on a single GPU.
Stack
Python, PyTorch, PyTorch Lightning, Hydra/OmegaConf, scikit-learn.
Status
Complete — trained checkpoint, evaluation pipeline, and visualized inference (EDC curve plots) all working.