Skip to content

v0.19.0

Latest

Choose a tag to compare

@hughcars hughcars released this 03 Nov 19:23
· 513 commits to main since this release
ca9a6a2

Palace 0.19 has been released! The full list of changes is in the CHANGELOG.md; these are the major updates. Palace now requires a C++20 compiler (CUDA 12 for GPU builds) and builds against MFEM 4.10.

Interior boundaries without mesh cracking — @simlapointe

Interior boundaries are no longer cracked by default. Impedance, rational impedance, and conductivity sheets are thin films coupling both sides by default, and "Crack": true models a thick conductor instead: two independent faces, each with the given surface impedance. These sheets were previously cracked by default, and each face of an impedance or rational impedance boundary instead had twice the given impedance, so that both faces together matched a single sheet. To keep the previous results, use "Crack": true with twice the impedance (twice "Rs" and "Ls", half "Cs"). Results move within the discretization error (about 3% on the CPW examples), and an interior SurfaceCurrent no longer decouples the two sides of its sheet. Error estimation uses broken finite element spaces, so discontinuous fields no longer need a modified mesh. PR 994, PR 1023.

Thin-film superconductors and inductance extraction — @dnpham23

Superconductor sheets extract the total (geometric + kinetic) inductance of thin films, and TwoSided films capture screening on both faces when the thickness is comparable to the penetration depth. Flux loops use the port excitation convention (FluxLoop[].Excitation) and renamed keys (FilmAttributes, PecPenetrationDepth). The FluxLoop hole inductance was low in 0.18 (2.00 → 2.85 pH on the circular_hole example); re-run existing cases. PR 929, PR 979, PR 992, PR 1051.

Wave ports and boundary modes — @simlapointe and @fauquexa

Voltage and impedance line integrals are exact for the discrete field (NSamples is removed), modes of superconducting and lossy lines are found on both sides of the target (WavePort[].Target), and port modes no longer depend on the mesh partitioning. Im{Z} in port-Z.csv had the wrong sign, and the second-order absorbing boundary reflected more than the first-order one. Adaptive sweeps replace the 2D port eigenvalue solves with a per-port reduced model. PR 909, PR 1049, PR 1056, PR 1066, PR 1067, PR 1070.

Corrected results

The 2D magnetic energy and 2D error estimates were off by orders of magnitude; boundary terms were stamped on every attribute sharing a material entry; surface flux and charge in lossy dielectrics ignored the loss; and multi-GPU runs with GPU-aware MPI gave wrong results from a race in hypre. PR 1069, PR 962, PR 953, PR 1062.

Other improvements

The nonlinear eigensolver has its own iteration limit (NonlinearMaxIts), AMS no longer leaks memory at order 1, linear solvers accept an absolute tolerance, and matrix-free operators skip elements with zero coefficients. Nastran and mixed-order Gmsh meshes read correctly, driven-sweep restarts work again, CSV tables are appended instead of rewritten, and the electrostatic solver has a manufactured-solution verification. Dependencies are pinned to releases and reported by palace --version. PR 1036, PR 1038, PR 734, PR 987, PR 951, PR 1122, PR 944, PR 942, PR 901, PR 1028.

New Contributors

Full list of contributors

@benvial, @dnpham23, @fauquexa, @FolletLouis, @hughcars, @jpsimas, @laylagi, @nikosavola, @phdum-a, @Sbozzolo, @simlapointe, @tcpdump-it

Full Changelog: v0.18.1...v0.19.0