Antiphon

Urban active noise cancellation driven by a learned acoustics model: a neural network that predicts how sound propagates through street geometry (complex transfer functions, magnitude and phase) so the ANC controller no longer needs expensive per-site measurement campaigns. The ANC corridor installation is the proving ground; the acoustics foundation model API is the product.

Headline result (v1, simulation): across 10 street geometries the model was never trained on, an FxLMS controller driven purely by model-predicted speaker-to-microphone paths achieves 90% of the noise cancellation that exactly-measured paths achieve, reaching full (>= 20 dB) cancellation in 15 of 20 cases and outright beating measured paths in several. The science: docs/science.md. The numbers: docs/results.md.

Third Axis AI Consulting / 316 Group.

What works today

• 2D FDTD acoustic solver (Yee staggered grid, split-field PML, rigid and impedance building walls). Validated against exact 2D analytical solutions: max 0.13 dB open-field amplitude error, single-wall interference within 1 dB.
• Facade materials (concrete/glass/brick absorption to impedance) and broadband noise sources (traffic, HVAC, construction) with octave-band metrics.
• Multi-channel FxLMS controller with Eriksson online secondary-path identification: 40+ dB tone reduction on FDTD-measured street-canyon paths.
• Synthetic data generator: randomized street canyons to complex transfer functions H(f) at 64 frequencies (30-430 Hz), HDF5, fully seeded.
• Foundation model (9.4M-param PyTorch: CNN geometry encoder + cross-attention queries) trained on 12,000 synthetic scenes; beats physics and statistical baselines by 2.4x on unseen geometries.
• Closed loop proven: FxLMS driven by model-predicted secondary paths achieves 90% of measured-path cancellation (capped metric) across 10 held-out scenes, beating measured paths outright in several. Full numbers: docs/results.md.
• Fast analytical solver (Green's functions + image sources) for interactive demos, refactored from the original prototype.

Current state: v1 pipeline complete; all success criteria met. See docs/results.md for the full evaluation and docs/STATUS.md for history.

Layout

src/antiphon/
├── simulation/     # Acoustic solvers and physics
│   ├── geometry.py     # UrbanGeometry, constants (y=0 = street centerline)
│   ├── fdtd.py         # FDTD solver (ground truth)
│   ├── analytical.py   # Green's function solver (fast demos)
│   ├── materials.py    # Absorption -> impedance
│   ├── sources.py      # NoiseSource, SpeakerArray, broadband generators
│   └── metrics.py      # Quiet zone, octave-band levels
├── anc/
│   └── fxlms.py        # Multi-channel FxLMS + online secondary-path ID
├── model/
│   ├── dataset.py      # Scene randomization + HDF5 dataset
│   ├── architecture.py # AcousticsModelV1 (10.4M params)
│   ├── train.py        # Training loop, baselines, scene-level splits
│   └── inference.py    # H prediction, sparse-H -> FIR filters
└── viz/                # Field plots and performance charts
scripts/            # CLI entry points (simulation, data gen, training, eval)
tests/              # 43 tests incl. FDTD-vs-analytical validation
refs/               # Original handoff material (do not modify)
docs/               # Proposal, status, figures

Quickstart

uv sync
uv run pytest                                     # full test suite
uv run python scripts/run_simulation.py           # 3-panel ANC comparison
uv run python scripts/run_simulation.py --sweep   # frequency sweep

# Full pipeline (compute-heavy; use a machine you can saturate)
uv run python scripts/generate_training_data.py --scenes 4000 --workers 6
uv run python scripts/train_model.py --data data/synthetic/train.h5
uv run python scripts/evaluate_closed_loop.py --ckpt data/runs/v1/best.pt

Figures are written to docs/figures/.

Roadmap

See refs/URBAN_ANC_HANDOFF.md for the full engineering plan and docs/STATUS.md for current progress:

1. Restructure reference script into this package (done, parity-tested)
2. 2D FDTD wave solver (Yee grid, PML boundaries) (done, validated)
3. Material absorption + broadband noise sources (done)
4. Multi-channel FxLMS controller (done)
5. Synthetic training data generation (done: 12k scenes / 288k samples)
6. Foundation model training (done: v2, val MSE 0.755 vs 1.85 baseline)
7. Closed-loop evaluation (done: 90% of measured-path performance)
8. Results report + figures (done: docs/results.md)
9. Next: broadband closed loop, 3D, web demo / investor materials