Project STAR · Intualex · Interactive method demonstrator

Physics-Constrained Learning — Thermoacoustic Stability Screening

Locating an instability boundary from few noisy evaluations · generic combustor
What this is — and is not. An interactive demonstration of the method on a generic textbook combustor (a one-dimensional acoustic network closed by a Crocco n–τ flame model). It shows how a physics-constrained fit locates a thermoacoustic instability boundary from very few noisy samples, where an unconstrained fit does not. It uses only generic, published parameters. It does not represent the STAR / LOUZEIRO engine and discloses no proprietary geometry, calibration or CFD. Feeding it any data only exercises the generic screening method. · Demonstração do método sobre um combustor genérico — não representa o motor nem revela qualquer parâmetro proprietário. · Companion to the paper: DOI 10.5281/zenodo.21923813.

Controls

Passive quarter-wave modes
f₁, f₂, f₃
Use your own data
Paste points as n,value (one per line). The two fits are applied to your points — the method is general and reveals nothing proprietary.

Stability screening — growth indicator vs flame gain n

true (physics) CPL (physics-constrained, linear) physics-free (cubic) true boundary ncrit ● noisy evaluations
2.00
true n_crit (generic model)
2.02 [1.78, 2.27]
CPL — median & 95% interval (200 runs)
2.00 [1.2, 2.8]
physics-free — median & 95% interval
How to read it. The physics says the modal growth rate is linear in the flame gain n near the boundary. The CPL fit builds that constraint in and recovers the marginal-stability gain ncrit tightly from a handful of noisy points; the physics-free fit (a cubic through the same points) overfits the noise and scatters — often failing to find a sensible boundary. That is the whole idea: a cheap, physically consistent pre-screen that tells you which operating points deserve an expensive high-fidelity (LES/CFD) run.