Agreement with Heinert's paper for cylindrical TR noise has now been achieved. Using the stationary state assumption to calculate the temperature profile, the computation time was reduced compared to the previous time-dependent approach. Here are the plots showing the agreement. I have shown the plots for a 1D axisymmetric model, in addition to a full 3D model in COMSOL. Both give the same result.
What went wrong? In the 1D axisymmetric case, it turns out that COMSOL has the incorrect cylindrical coordinate Laplacian for the coefficient form PDE interface. I corrected for this by expanding the Laplacian with the product rule, giving an extra -kappa/r ''convection coefficient'' in the PDE interface. Meanwhile, the 3D case worked from the beginning, since its Laplacian was in cartesian coordinates.
The next goal is to use a stationary state method to calculate TE noise for a test case.