Tutorial · XD² Lab · University of Tennessee, Knoxville
A supersonic intake slows Mach 2 air through a cone shock and a terminal shock near the cowl lip. Close the throttle at the back of the duct far enough and the terminal shock stops holding still. It is pushed out onto the cone, the duct empties, the shock is swallowed, the duct refills, about a hundred times a second. This is inlet buzz. Our current work treats its onset as a Hopf bifurcation of the discretized flow equations and finds it from one eigenvalue problem instead of a long time-accurate simulation. This page builds the idea up with a two-variable model you can solve by hand, then maps each step back to the intake.
After spatial discretization, the flow equations become a large system of ordinary differential equations, \(\dot w = f(w;\,\mathrm{TR},\,x)\), with the throttle ratio TR as the operating parameter and the cowl shape \(x\) as the design. For the intake, \(w\) holds 347,904 numbers: six per cell. Every idea we need already shows up in a system with two:
This model comes from our earlier paper on the stability of Hopf bifurcations (He et al., 2025), where it served to verify derivatives and to demonstrate an optimization. Here \(\mu\) is the bifurcation parameter and plays the role of throttle closure: raising \(\mu\) pushes the system toward instability, as closing the throttle does. The two design variables are \(x = (x_1, x_2)\), and \(\beta = -0.1\) is a small constant load that moves the equilibrium off the origin, as in the computations of that paper.
| Two-variable model | Nagashima intake (RANS) | |
|---|---|---|
| State \(w\) | 2 numbers | 347,904 numbers (6 per cell) |
| Parameter | \(\mu\), unstable as it increases | throttle ratio TR, unstable as it decreases |
| Design \(x\) | \(x_1, x_2\) | 12 cowl-shape variables |
| Equilibrium | \(f(w_0) = 0\) by Newton | steady RANS by Newton–Krylov |
| Linear operator | 2 × 2 matrix \(A\) | Jacobian with 26.6 million nonzeros |
| Onset | \(\mu_c\) | TRc = 0.86–0.88 |
| Frequency at onset | \(\omega_c \approx 0.96\) rad per unit time | 112–115 Hz |
The process has three steps.
Try both designs. With the optimized design, crossing \(\mu_c\) gives a small, stable oscillation that grows smoothly as you go further. That is a supercritical Hopf bifurcation, and it is what the intake does: its RANS cycles appear at TRc with a small amplitude that grows as the throttle closes.
With the baseline design the crossing is subcritical. Below \(\mu_c\) the equilibrium survives small kicks, but an unstable cycle (dashed) surrounds it and a kick that lands outside escapes. Past \(\mu_c\) there is nothing nearby to settle on and every trajectory runs away. In a machine this shows up as a sudden jump to large oscillations that does not go away when you back the parameter off a little.
A bifurcation diagram puts the long-time behavior at every parameter value into one picture. The horizontal axis is \(\mu\). The vertical axis is the peak-to-peak amplitude of \(w_1\) after the transients are gone. The equilibrium has zero amplitude and is drawn solid where it is stable, dashed where it is not. The cycles come from time integration started near the cycle: forward in time for a stable cycle and backward in time for an unstable one, which makes an unstable cycle attract.
Near the Hopf point the dynamics reduce to one complex amplitude \(z\), with \(w \approx w_0 + 2\,\mathrm{Re}(z\,q)\) and
The growth rate \(\sigma\) passes through zero linearly at \(\mu_c\), so the squared amplitude of the cycle grows linearly too: \(A^2 \propto |\mu - \mu_c|\). The first Lyapunov coefficient \(l_1 = \mathrm{Re}\,c_1/\omega\) decides the side. With \(l_1 < 0\) stable cycles appear past \(\mu_c\) (supercritical); with \(l_1 > 0\) unstable cycles exist before it (subcritical). Computing \(l_1\) takes only the eigenvectors and the second and third derivatives of \(f\) at the Hopf point. The dotted curve below is that local prediction, and the dots come from time integration.
The intake study does not compute \(l_1\). It checks the signatures of a Hopf bifurcation directly. A time-accurate run at TR = 0.90 decays at \(\sigma = -129\) 1/s and 121 Hz, against the eigenvalue \(-160\) 1/s at 112.5 Hz. The squared plenum swings of the cycles at TR = 0.88, 0.87, and 0.85 lie on a straight line that reaches zero at TRc = 0.881, and their frequency, 114 Hz, matches the eigenvalue frequency at the crossing. Switch the axis to A² to see the straight line in both columns. Past TR = 0.80 the intake leaves the local law and jumps to large cycles, a reminder that the normal form only describes the neighborhood of the onset.
The eigenvalue depends on the equilibrium, because \(A\) is evaluated at \(w_0\). The equilibrium does not depend on the eigenvalue. That one-way dependence splits the analysis into two levels.
Level 1 (bottom). Solve \(f(w_0;\,\mu, x) = 0\) for the base flow.
Level 2 (top). With \(w_0\) fixed, solve for the eigenpair \(v = (q, \lambda)\):
The last row fixes the length and phase of \(q\), which an eigenvector does not have on its own. Here \(c = e_1\), so \(q_1 = 1\). Stacked together, the Jacobian of both levels with respect to \((w_0, v)\) is block lower triangular. The forward solve runs bottom to top. The adjoint solve, which we need for derivatives, uses the transpose, which is upper triangular, so it runs top to bottom.
\(\hat r(v;\, w_0, x) = 0\) gives \(\lambda = \sigma + i\omega\)
\(\big(\partial \hat r/\partial v\big)^{\mathsf T}\hat\psi = e_\lambda\)
\(f(w_0;\,\mu, x) = 0\), Newton
\(A^{\mathsf T}\psi = -\,\mathrm{Re}\big[(\partial \hat r/\partial w_0)^{\mathsf T}\hat\psi\big]\)
Columns: \(w_0\) (n of them), then \(v\) (n + 1). Here n = 2. For the intake n = 347,904, and the top level is a complex system of 347,905 unknowns.
| k | w₁ | w₂ | ‖f‖ |
|---|
For the intake, Level 1 is the steady RANS solve with Newton–Krylov iterations, about 4 core-hours where it converges. Level 2 is a shift-and-invert Krylov–Schur solve with the exact Jacobian, which the flow solver already assembles for its adjoint: 30–46 s on 16 cores. Past the onset no steady solver reaches the unstable equilibrium, so Level 1 is the hard part, a limitation the study states openly.
Design asks how the growth rate changes with the shape: \(d\sigma/dx\). Finite differences need one full two-level solve per design variable, and for the intake each one starts with a steady RANS run. The adjoint method gives every component of the gradient for the cost of two linear solves.
A design change acts through two paths. It changes \(A\) directly with \(w_0\) held fixed. It also moves the base flow, and \(A\) changes because it is now evaluated somewhere else. The chain rule adds the two paths, and the adjoint evaluates both without ever forming \(dw_0/dx\):
Solve for \(\hat\psi\) first, then \(\psi\), in the order the block-triangular structure sets. Two facts make this cheaper than it looks. The top-level adjoint \(\hat\psi\) turns out to be the left eigenvector of \(A\), so the first term is the textbook eigenvalue derivative \(y^{\mathsf T}(\partial A/\partial x_k)\,q \,/\, y^{\mathsf T}q\). The bottom-level system has the same matrix as a steady flow adjoint, with a right-hand side built from the second derivative of \(f\) in the direction \(q\).
| Level 2 term | Level 1 term | adjoint total | finite diff. |
|---|
Central differences with step 10⁻⁶ agree with the full adjoint to about 10⁻⁹, their own truncation error. Drag β to 0: the equilibrium returns to the origin, the Level 1 term vanishes, and both bars drop to the floor. With β ≠ 0, skipping Level 1 gives a gradient that is off in the third or fourth digit here. For a flow whose shock moves with the design, there is no reason to expect the error to stay that small.
The same machinery gives \(d\sigma/d\mu\), and with it a way to find the onset without a sweep. One Newton step on \(\sigma(\mu) = 0\),
moves from a single stable base flow straight toward the Hopf point. On the intake, the slope between the equilibria at TR = 0.97 and 0.90 is \(-3.9\times 10^{3}\) 1/s per unit TR; one step from \(\sigma = -160\) 1/s at TR = 0.90 lands at TR ≈ 0.86. The adjoint supplies the slope from one base flow instead of two.
| step | μ | σ | dσ/dμ |
|---|
In the earlier paper the model was optimized to push the onset to larger \(\mu\) while keeping it supercritical:
The constraint matters. The baseline \(x = (1, 1)\) has a late onset, \(\mu_c = 0.985\), but it is subcritical (\(l_1 = 1.50\)) and fails without warning. The optimizer gives up some onset for a safe bifurcation and stops at \(x = (1, 0.436)\) with \(\mu_c = 0.720\) and \(l_1\) on its bound.
Drag the design point. The color shows \(\mu_c\). The red line \(l_1 = 0\) separates subcritical from supercritical designs, and the hatched region violates the constraint. The arrows are the gradients: \(\nabla\mu_c\) comes from the two-level adjoint of Section 4 through \(d\mu_c/dx = -(d\sigma/dx)/(d\sigma/d\mu)\) at the Hopf point, and \(\nabla l_1\) is computed here by finite differences (the paper derives its adjoint).
Amplitude against μ − μc. Supercritical designs grow a stable cycle to the right; subcritical designs carry an unstable cycle to the left.
| it | x₁ | x₂ | μc | l₁ |
|---|
A small sequential quadratic programming loop: linearize μc and l₁, take the best step inside the box and the linearized constraint, repeat.
The intake version of this problem is set up the same way in our study: the buzz-mode growth rate at TR = 0.90 as the objective, 12 cowl-shape variables near the lip where the mode peaks, a constraint that keeps the total-pressure recovery at its baseline, and gradients from the two-level adjoint. The size changes from 2 unknowns to 347,904; the structure of the calculation does not.