WaveBench

Laboratory-grade 1D engine gas-dynamics, acoustics and forced-induction design suite.

• Platform: .NET 10 (LTS) · Windows desktop (WPF) · headless CLI on any .NET target
• Licence: Apache-2.0
• Audience: Formula SAE teams, race engine developers, professional engine designers, DIY engine enthusiasts
• Scope: intake and exhaust wave tuning · camshaft timing · collector configuration and cylinder pairing · exhaust sound design and auralisation · turbocharging and supercharging · multi-objective optimisation

No telemetry. No network calls at runtime. Your designs, dyno data and audio never leave your machine.

---

Fifteen minutes to a torque curve

dotnet build
dotnet run --project src/WaveBench.Cli -- sweep examples/single-360.json \
           --from 4000 --to 9000 --step 500 --plot sweep.png

That is a converged torque curve. To check it is a property of the engine rather than of the grid it was solved on:

dotnet run --project src/WaveBench.Cli -- mesh examples/single-360.json --rpm 7000

For the desktop app, open a template, describe your engine in Design, press Run, and read Results. The full walkthrough is docs/user-guide.md.

---

Status

25 of 26 phases complete. 1016 tests green (608 unit + 408 verification), none skipped. The complete build specification lives in docs/WaveBench-Master-Plan.md — a staged build contract with 26 phases, each with a hard acceptance gate. Physics before pixels: Phases 0–15 produced a headless, test-covered engine; no UI existed before Phase 16.

What works today:

• Gas dynamics — species-resolved 1D MUSCL-Hancock + HLLC, verified against exact Riemann solutions; well-balanced variable area; friction, heat and wall-thermal sources; reservoir, orifice, plenum and junction components; an FSAE restrictor that chokes at theory.
• Engines — motored and fired single- and multi-cylinder engines with wave-tuned VE curves, Wiebe combustion, knock tracking, per-cylinder breakdowns.
• Acoustics — transfer-matrix engine cross-validated against the nonlinear solver to 0.45 dB, collector pulse-timing analysis that reproduces the crossplane-vs-flat-plane signature from firing order alone, and audio synthesis with phase-coherent crank-angle wavetables and BS.1770 level-matched A/B.
• Forced induction — compressor and turbine maps with a digitiser, steady matching, coupled unsteady flow, transient spool with a sensitivity band, Greitzer surge, and a Boost workspace that draws the operating line on the map.
• Optimisation — DOE, Morris screening, CMA-ES, NSGA-II, Bayesian optimisation with a Gaussian-process surrogate, Nelder–Mead and Powell refiners, a surrogate inner loop, and a Pareto explorer with click-to-audition.
• The learn layer — why-text and a typical range on every field, "Show me" parametric sweeps, concept explainers, guided tours and a generic-defaults banner.
• Reporting — one-click PDF and HTML, both rendered from one document so they cannot disagree.

Remaining: Phase 11's four psychoacoustic metrics, and Phase 15's transient-spool validation against a measured case. Both are deferred for stated reasons rather than skipped — see the gallery.

---

Headless CLI

wavebench info    examples/single-360.json
wavebench run     examples/single-360.json --rpm 5000
wavebench sweep   examples/single-360.json --from 4000 --to 9000 --step 500 \
                  --db results.db --plot sweep.png
wavebench mesh    examples/single-360.json --rpm 7000
wavebench render  examples/single-360.json --from 2500 --to 7500 --seconds 9
wavebench report  examples/single-360.json --from 4000 --to 9000 --step 500 --out reports
wavebench validate --out validation

report solves the sweep, runs the mesh-sensitivity study and writes report.html and report.pdf: the complete model dump with the origin of every value, the geometry, every figure, the convergence evidence, the acoustics, the turbo match, every assumption with its citation, and a validation statement. It is built to be handed to an FSAE design-event judge, which is why it states its own weaknesses rather than leaving them to be found.

render solves an rpm grid, builds crank-angle wavetables from the solved pressure history and synthesises phase-coherent audio — 24-bit/48 kHz WAV with separate exhaust/intake stems and a provenance sidecar recording the model hash, seed and resolved bandwidth. Content above that bandwidth is labelled as not physically resolved rather than presented as prediction.

---

Validation gallery

Every claim is backed by something outside this tool, or it is listed below as not backed. The full citation list — every correlation, its source and its validity range — is docs/citations.md.

Published-data case

The open-access CSU thesis runner-length study. WaveBench reproduces the published optimum exactly at 800 mm and within the 250 rpm gate at 600 mm:

The short-runner discrepancy is documented rather than tuned away — validation/yin-runner-length.md and docs/physics.md §1.9.

Analytical anchors

These have exact answers, which makes them better anchors than any measurement. All run on every commit.

Case	Result
Shock tube against the exact Riemann solution	density, velocity and pressure match
Stationary taper, well-balancedness	spurious velocity < 1e-10 m/s
Steady pipe friction vs Darcy–Weisbach	within 1%
Steady wall heat transfer vs analytical	within 1%
Isentropic nozzle through to choking	matches compressible-flow relations
Organ-pipe resonance of a closed-open duct	lands on the analytical n·c/4L series
Junction under a pulse of 69% of mean pressure	0.07% error
TMM against the nonlinear solver	0.45 dB
FSAE restrictor	chokes at theory

Known-answer optimisation benchmarks

Problem	Result
Sphere 4-D, 20 seeds	20/20 to 1e-8
Rosenbrock 4-D, 20 seeds	19/20 to 1e-6
Rastrigin 2-D, 20 seeds	15/20 find the global optimum
Ishigami Sobol indices	match the analytic values
ZDT1 / ZDT2 fronts	match the known fronts

Not validated

Listed because a project that only publishes what works is not publishing evidence.

Gap	Why	What to do about it
Transient spool vs a measured case	No redistributable measured dataset was found (docs/physics.md §6.4 records what was checked)	Self-consistency is checked instead; compare against your own logged data before quoting a spool time
Measured order levels	No redistributable dataset	Acoustic comparisons are relative between designs, not absolute
Diabatic correction vs a measured on-engine outlet temperature	No redistributable dataset	—
ISO 532-3, ECMA-418-2, fluctuation strength, DIN 45681	Each needs verification against reference signals that cannot be redistributed	Character is reported through the metrics that are verified

Have dyno data with known geometry — especially FSAE? Please open an issue. A measured case with provenance is the most valuable contribution this project can receive.

---

Installing

The release is a per-user MSI: it installs into %LocalAppData%\Programs\WaveBench without elevation, adds a Start Menu entry, and carries both the desktop app and the wavebench CLI. It is not signed — Windows will name an unknown publisher at a SmartScreen prompt and let you decide. Uninstall removes everything it added.

To build the installer yourself, see packaging/.

Building

dotnet build
dotnet test

Requires the .NET 10 SDK. Everything except WaveBench.App is a cross-buildable class library or the CLI; the desktop app is WPF and therefore Windows-only.

Solution layout

Project	Purpose
WaveBench.Core	Physics: thermodynamics, 1D solver, components, engine model (no UI, no I/O beyond streams)
WaveBench.Acoustics	TMM, radiation, order analysis, psychoacoustics, synthesis
WaveBench.Boost	Turbo/supercharger maps, shaft dynamics, thermal states, boost control
WaveBench.Model	Serialisable model tree, strongly-typed units, validation rules, provenance
WaveBench.Analysis	Post-processing, FFT, wave decomposition
WaveBench.Optimize	DOE, optimisers, surrogates, constraints
WaveBench.ViewModels	All UI logic and the report generator — plain net10.0, zero UI types
WaveBench.Cli	Headless runner and scripting entry point
WaveBench.App	WPF desktop app

WaveBench.ViewModels holding every piece of UI logic with no UI types in it is what lets the CLI generate the same report the app does, and lets three quarters of the interface be tested without a window.

Tests: WaveBench.Core.Tests (unit), WaveBench.Verification (§6.1, per-PR CI), WaveBench.Validation (§6.2, nightly), WaveBench.Bench (BenchmarkDotNet).

Documentation

Document	What is in it
docs/user-guide.md	How to use it, start to finish
docs/citations.md	Every correlation, its source and its validity range
docs/physics.md	What the models do and the measurements behind them
docs/acoustics.md	The acoustic engine
docs/numerics.md	Scheme behaviour, bandwidth, mesh guidance
docs/WaveBench-Master-Plan.md	The build contract
CHANGELOG.md	Every phase, with its measurements

Ground rules (from the plan, Part 0)

1. Do not skip phases; every gate must pass before proceeding.
2. TDD is mandatory in the physics layers, tested against analytical or published references.
3. Every empirical correlation is cited in an XML doc comment with its validity range.
4. WaveBench.Core never references a UI assembly (enforced by an architecture test).
5. Determinism: same input file → bit-identical results.
6. Docs ship in the same commit as the code.