Skip to content

Shape optimization

In lumped-element experiments, changing a detector's dimensions changes its signal coupling, inductance and stored energy. In this guide we compute shape sensitivities and run a bounded search. A shape sensitivity measures the change in one of these quantities per change in a design parameter. axionMultiPhysics provides optimizers that use those derivatives to search for a local maximum of sensitivity or a local minimum of stored energy within chosen bounds.

For an objective \(J\), parameter \(p\), field unknowns \(y\) and residual \(R\), the discrete adjoint gives

\[ \frac{dJ}{dp} = \frac{\partial J}{\partial p} + \lambda\cdot\frac{\partial R}{\partial p}, \qquad S\lambda = -\frac{\partial J}{\partial y}. \]

Here \(S\) is the symmetric tangent operator and \(\lambda\) is the adjoint field. The mesh moves with the geometry while keeping its connectivity. The adjoint supports axisymmetric magnetoquasistatic designs with linear materials; 3D, full-wave results and RF loss budgets are not differentiated.

For a solenoid, x is the coax inner radius divided by the winding inner radius. With SciPy installed, we compute derivatives of the scan figure of merit to quantify its dependence on x:

import axionmultiphysics as axmp

design = axmp.template("solenoid")
result = design.solve()
s = result.sensitivities(params=["x"], objectives=["F"])
print(s["gradient"]["F"]["x"], s["relative"]["F"]["x"])

The first number is the derivative in SI units; the second is the fractional change in the objective per fractional change in x. These are derivatives on the current mesh: mesh convergence should be established before optimizing.

We now run a bounded search for a local minimum of the stored energy the solenoid needs to reach the scan requirement:

opt = axmp.optimize(design, objective="W_at_req:optimistic",
                  params={"x": (0.3, 0.8)}, max_iter=30)
print(opt["converged"], opt["message"])
print(opt["parameters"], opt["value"])

The value is in joules, after scaling the field to the selected scan requirement (the optimistic benchmark is an illustrative preset, and the default scan target an illustrative choice). axmp.objectives() lists the other objectives, among them the coupling ("F" with sense="max") and the peak fields at the requirement. converged and message describe the search termination; design contains the returned design document and history records the iterations. Field caps are only reported unless added as explicit constraints on sc_surface_at_req or B_peak_at_req; sharp superconductor corners should be rounded before optimizing a peak-field constraint.

For sensitivities involving nonlinear iron or changes to mesh connectivity, we pass method="fd_remesh" to result.sensitivities. The API reference lists arguments and return values.