CST Studio Suite 2024 FIT/GPU: temporal qualification of small differential S-parameters and late energy decay

 

Dear SIMULIA CST Support,

We request guidance on a controlled numerical qualification problem in a 3D microwave imaging study. We need to establish the accuracy of DeltaS = S_phantom - S_background, rather than relying only on convergence of absolute S-parameters.

Configuration: CST Studio Suite2024, time-domain FIT, GPU acceleration, 55,182,739 cells with a0.45mm local mesh limit, matched background/phantom axes, Tx4,50ohm reference impedance,0.5-2GHz solver band, evaluation at0.6/0.8/1.0/1.2/1.4GHz, one E-field monitor at1.2GHz,20pulse-width maximum. ConvPML and automatic stabilization were retained. GPU is reported as unverified; the hardware override was used and is disclosed, not assumed validated. Windows host, Intel Core Ultra9 275HX, approximately32GB RAM. The exact service pack and GPU model can be supplied after verification.

Observed results:
- At fixed mesh, changing steady-state energy limit from-60 to-80dB changed matched DeltaS by12.911%,3.891%,2.989%,2.644%,2.760%. Absolute S and E1.2 were much less sensitive.
- A background-only-100dB attempt reached the20pulse cap near94.77ns and-89.97dB. The energy criterion failed; this result is excluded and no corresponding phantom attempt was launched.
- At-80dB, a matched pair changing only estimated PML reflection from1e-4 to1e-6 changed DeltaS by0.718%,0.077%,0.146%,0.032%,0.041%. It did not establish temporal convergence or the cause of the late energy floor.
- Rebuilding S from saved incident/outgoing port signals agrees with exported S within approximately8e-6 relative norm error. Removing the last2.5% of recorded signals changes600MHz DeltaS by approximately2.31% for the new pair. This is within-record sensitivity, not a bound on the unrecorded future tail.
- The subtraction amplification indicator (norm(Sphantom)+norm(Sbackground))/norm(DeltaS) is approximately269 at600MHz. This does not identify floating-point roundoff as the cause.

Questions:
1. Which supported diagnostics distinguish physical long-lived energy from numerical residual energy in this exact FIT/GPU implementation?
2. Are there applicable precision controls for this solver, and how should we verify effective precision? We found precision controls for a different frequency-domain solver and have deliberately not applied them here.
3. Would one reduced TimeStepStabilityFactor test be appropriate despite no demonstrated energy growth or instability? If so, which factor and comparison would be diagnostically meaningful while holding geometry, materials, mesh, physical excitation and observation policy fixed?
4. What is the recommended method to demonstrate temporal accuracy of a small difference between two simulations when a global energy threshold and stable absolute S are insufficient?

A geometry short-circuit notice remains unresolved and is disclosed. We are not requesting a new PBA investigation in this case; please focus on the temporal accuracy and differential-signal questions above, and flag if a conclusion cannot be made within that scope.

Acceptance remains relative2% on DeltaS itself and maximum absolute complex difference0.01, with no phase alignment. We do not want to accept failed solves, relax criteria, or extend time windows without a justified protocol. Please recommend one discriminating test before a further expensive run.

Attached: CST_temporal_review_evidence.zip (62 source evidence files and SHA256 manifest). Additional local review: B_DECISIVE_DIFFERENTIAL_AUDIT.json. Full project files have not been included in this initial evidence package and can be considered separately if needed.

Kind regards,