Tutorial validation record#

All six backend workflows were executed locally on 14 September 2026 using the scripts included in this repository. This record describes the actual Windows runs used to produce the tutorial figures. A separate GitHub Actions run also built and executed all six workflows on Linux, Windows and macOS, as recorded below. The current spherical examples were recalculated after the harmonic cutoff audit, and additional QSEIS regional workflows were run at 300/600/900 km.

Fresh 2 Hz displacement and stress comparison#

On 15 September 2026, all five dynamic backends and two Gaussian smoothing controls completed new calculations at source depth 10 km, receiver depth 1 km and distances 300/600/900 km. The shared AK135-FC model, 0.25 s sampling, 2 Hz maximum and 1.25 s sin² source differ from the lightweight tutorials below. See the fresh report and public validation summary.

Calculation	Complete processed displacement shape	Runtime	Recorded output size, decimal
QSEIS06, native 0.05 smoothing	`(3, 3, 2048)`	971 s	2.3 MB
QSEIS2025, 0.05 smoothing control	`(3, 3, 2048)`	983 s	5.0 MB
QSEIS06, isolated point-source control, eight frequency tasks	`(3, 3, 2048)`	609 s	24.9 MB
QSEIS2025, point source, eight frequency tasks	`(3, 3, 2048)`	689 s	52.8 MB
SPGRN2012	`(3, 3, 2048)`	1419 s	1.207 GB
SPGRN2020	`(3, 3, 2048)`	1542 s	2.777 GB
QSSP2020, new spectra and six basis-source syntheses	`(3, 3, 2048)`	1908 s	11.715 GB

QSEIS2025 and QSSP2020 stress arrays have shape (3, 6, 2048) and units Pa. The public comparison arrays contain 1601 samples on 0–400 s. Complete native records, finite values, source area/centroid/spectrum, executable hashes and QSSP’s native rate/direct-output integration identity were checked. All 636 published metric rows were independently recomputed from the public arrays, with a maximum discrepancy of 1.42e-14 percentage points.

Matched-radius QSEIS versions were bit-identical in displacement at both ratios 0.05 and 0. The latter QSEIS06 result requires an isolated control build; it is not the released default. The isolated QSEIS builds increase layer capacity to 2048, and the point calculations sum eight disjoint frequency ranges with validated linear superposition. Their source diffs and executable hashes are included in the public data record. No installed solver or package default was changed by these controls.

These are Windows 11 measurements using Python 3.12.13, NumPy 2.3.5, SciPy 1.18.0, Matplotlib 3.11.0 and ObsPy 1.5.0. Some calculations overlapped, so their elapsed times are not a controlled performance benchmark. Sizes include each result directory’s retained data and exclude later website assets. These scientific runs were not added to per-PR tutorial CI and do not extend the older cross-platform validation claim to this new setup.

The portable plotting script replots the committed comparison arrays with NumPy and Matplotlib. It does not recompute a native Green’s library.

Environment and results for the lightweight tutorials#

The local environment was Windows 11 build 26200, 64-bit CPython 3.12.13, NumPy 2.3.5, SciPy 1.18.0, pandas 3.0.0, Matplotlib 3.11.0 and ObsPy 1.5.0. Native executables were installed in the same Conda environment. Commands used conda run -n ffipy python examples/<script>.py.

Times below measure preparation, solver execution, output reading, checks and plotting after imports; they exclude interpreter startup and package installation. Sizes include the retained solver ASCII files, binary files, spectra and figures. They are measurements for these small examples, not performance guarantees.

Workflow	Command argument	Validated output shape	Time	Output size	Machine-readable record
QSEIS2025 introduction	`--observables all`	displacement `(3, 3, 201)`; strain/stress `(3, 6, 201)`	13.3 s	1.21 MiB	JSON
QSEIS06 introduction	default	`(3, 3, 256)`	10.0 s	0.38 MiB	JSON
SPGRN2012	default, impulse + source-matched library	`(3, 3, 256)`	40.5 s	203.39 MiB	JSON
SPGRN2020	default, complete wavefield	`(3, 3, 256)`	62.5 s	266.14 MiB	JSON
QSSP2020	default, harmonics 2000/8000	`(3, 3, 256)`	86.3 s	657.98 MiB	JSON
EDGRN2 → EDCMP2	default, including both solvers	`(3, 3)`	3.1 s	0.17 MiB	JSON
QSEIS2025 regional	`--regional --observables all`	displacement `(3, 3, 256)`; strain/stress `(3, 6, 256)`	175.9 s	3.66 MiB	JSON
QSEIS06 regional	`--regional`	`(3, 3, 256)`	175.1 s	1.19 MiB	JSON

The dynamic array axes are distance, component and sample. The static axes are distance and component. The default displacement-only QSEIS2025 command was also executed independently in a fresh output directory.

On 15–16 September 2026, the QSEIS06 introduction, both regional QSEIS rows and SPGRN2012 were rerun in fresh directories after the examples switched to requesting displacement, strain and stress directly from the readers. The QSEIS arrays matched the earlier explicit rate integration to within 6e-16 of each peak. The SPGRN2012 displacement differed by relative L2 2.3e-7, because the source-matched Green functions are stored as float32.

The QSEIS2025 run and figures were refreshed after cropping every exported observable to 0–100 s inclusive. At 0.5 s spacing, the NPZ arrays contain 201 samples; each was checked against the first 201 samples of the previous full waveform. Its native library still contains 256 samples over 127.5 s. The JSON records the exported interval in output_time_range_s separately from the native time_window_s. The refreshed displacement-only run took 11.4 s.

Matched frequency band, mechanism and temporal source#

All five regional scripts were run in fresh directories with 4 s sampling, 1024 native FFT samples, frequency spacing 1/4096 Hz and a 0.125 Hz maximum. Native spectrum headers (SPGRN/QSSP) and input/output grids (QSEIS) were checked: 512 nonnegative bins are computed through 511/4096 Hz, and Nyquist is zero. The three old spherical libraries were rejected because their native headers retained only 257 bins, even before checking the declared JSON settings.

Every script uses the same strike/dip/rake (30/45/90 degrees), azimuth (30 degrees), scalar moment (10^15 N m), source depth (10 km), receiver depth (0 km), and effective normalized 64 s sin-squared moment-rate pulse. SPGRN2012 now uses full-wavefield spectra and the native zero-duration impulse branch. Each complete 1024-point impulse Green function is convolved forward with the analytic source at the native complex frequencies and written to a source-matched library. Reading that library with output_type="disp" integrates once before the script crops to 256 samples. Quadrature independently verified the full-band source transform with relative L2 error 1.75e-15. Input, native-velocity, matched Green-function and NPZ hashes were unchanged by compatible reuse. The old positive-duration library was rejected.

QSEIS retains its compensated custom source. Its independent spectral check now covers 0–0.125 Hz. Both QSEIS native calculations completed normally; their displacement samples were identical and their exported tensor arrays were finite. QSEIS2025’s original short example still exports 0–100 s.

The earlier 64 s comparison uses the common native origin-time intervals [3,500], [40,500] and [78,500] s, interpolated to 1 s without amplitude or time-shift fitting. Relative ENU L2 differences from SPGRN2020 are:

Backend	300 km	600 km	900 km
SPGRN2012	1.2266%	0.0739%	0.0535%
QSSP2020	1.2772%	0.4085%	2.3022%
QSEIS06 / QSEIS2025, default spatial smoothing	12.7410%	19.5740%	22.7979%
QSEIS2025, point-source control	11.4926%	17.0758%	19.7176%

The point-source control only changes QSEIS2025’s Gaussian spatial-source ratio from 0.05 to zero; its model and temporal source samples are unchanged. That additional Windows run took 615.795 s and retained 1,005,168 bytes in 27 files; its (3,3,256) displacement was finite. See the point-source record. It reduces only part of the residual and was not added to every CI calculation. QSEIS06 has the ratio fixed at 0.05 in the solver. No Fortran algorithm was changed.

The fresh runtimes and output sizes above are preserved separately from the short compatible rereads used to check the final scripts and metadata. Archived JSON records identify those post-run checks explicitly. Older frequency-cutoff, harmonic and STF comparisons remain in the historical audit. The initial cross-platform archive below predates these updated settings.

Initial cross-platform installation and execution#

GitHub Actions run 34801126315 built commit 0d38d44 before the QSEIS2025 100 s crop. These archived records therefore retain its original 256-sample exported waveforms; the current 201-sample exports are verified by the refreshed local run above. The CI run installed source in fresh Conda environments, using Python 3.12.14 on Linux x86-64, Windows x86-64 and macOS arm64. All six workflows passed on every platform. The Windows job used gfortran=15.2.0=hf1b5d6d_19; the installation guide records why that exact compiler build is selected.

Each cell below gives calculation time / retained output size. The timing boundary is the same as the local table; compiler installation and compilation are excluded. QSEIS2025 includes displacement, strain and stress.

Workflow	Linux	Windows	macOS
QSEIS2025	15.4 s / 1.19 MiB	23.2 s / 1.21 MiB	9.1 s / 1.19 MiB
QSEIS06	14.2 s / 0.37 MiB	19.9 s / 0.37 MiB	8.3 s / 0.37 MiB
SPGRN2012	17.3 s / 12.54 MiB	18.5 s / 12.54 MiB	9.8 s / 12.54 MiB
SPGRN2020	17.4 s / 17.09 MiB	19.2 s / 17.10 MiB	9.7 s / 17.09 MiB
QSSP2020	25.4 s / 63.72 MiB	27.9 s / 63.73 MiB	19.1 s / 63.72 MiB
EDGRN2 → EDCMP2	0.3 s / 0.16 MiB	1.6 s / 0.17 MiB	0.3 s / 0.16 MiB

The preserved CI records include all 18 run summaries, exact platform/Python/dependency versions, dimensions, component names, units, amplitudes and output sizes. The 24 uploaded NPZ arrays were reopened and checked for finite values, matching shapes, component labels and units. The workflow also retains plots and selected native inputs as downloadable artifacts for 14 days; the JSON records here remain part of the documentation after those artifacts expire.

These checks exercise editable source installation. They do not claim a standard-wheel validation, multi-node MPI validation or scientific convergence for arbitrary models. They also do not exercise every Python version in the package support range.

Documentation checks#

The isolated Python 3.12 documentation environment uses the locked requirements in this repository. Sphinx’s strict HTML build passed without warnings; the explicit API checker covered 92 functions, classes and methods. Internal files and anchors, MathJax formulas, figure loading, search, and desktop (1440 px) and mobile (390 px) layouts were checked. All 25 modified Python source files had identical ASTs after removing docstrings.

The focused TauP suite passed 12 tests, with one installed-wheel-only test skipped in the source environment. A source-distribution archive was checked to include documentation, figures and executable tutorial sources, excluding generated HTML and calculation libraries.

What was checked#

Every dynamic example returned three-component displacement at three distances, with finite values and a nonzero waveform. The QSEIS2025 introduction exports 201 samples over 0–100 s; its regional variant and other dynamic examples export 256 samples. QSEIS2025 strain and stress have six components on the same exported distance/time grid as displacement in each mode.
EDGRN generated the layered kernels; EDCMP used those kernels for all five mechanism bases at both source depths. ASCII-to-binary conversion produced the bulk shape (2, 1, 5, 5, 3), and the three queried displacements were finite.
Component labels and time axes were checked against the readers. Vector figures use east, north, up; QSEIS2025 tensors use EE, EN, EU, NN, NU, UU. SPGRN2012 uses its distance-dependent native start time, and the revised SPGRN2020 example uses the origin-time starts stored in its native binary headers.
The static material lookup uses four-column noQ.nd, avoiding an invalid six-column material reshape. Its result is explicitly multiplied by seismic moment after EDCMP’s unit-moment normalization.
Figures were inspected for readable axes, legends, component labels and layout. The spherical examples use a longer spectral window than output window to reduce periodic end-of-window contamination.

Reproduce and interpret the record#

Run the commands in the backend tutorials. Each script writes its own summary.json, with the environment, elapsed time, array dimensions, maximum absolute value and output size. --reuse repeats the reader/plot checks without native recomputation and writes summary-reuse.json.

The base runs check installation and workflow; the additional comparison records verified frequency grids, temporal sources and specific numerical checks. The tutorial model uses the bundled AK135 elastic structure with explicit constant Qp = 600 and Qs = 300. Regional examples truncate it at 809.5 km; spherical examples retain the full structure. The figures are not a claim that these different discretizations, source time functions and solver approximations produce interchangeable research results. Check convergence in spectral window, bandwidth, spatial sampling, harmonic or wavenumber settings and model resolution for the scientific problem of interest.

EDGRN requires two source depths. Static queries are kept within the distance table because EDCMP finite source corners can cross a table edge. Multi-node MPI and large production calculations were not exercised by these local runs.