Choose mesh settings and measure memory and runtime

Before starting

The useful question is how much refinement your result needs and what that refinement costs on your machine. Start from a completed representative calculation, then change one setting at a time. A mesh count by itself cannot predict memory use or runtime for a different geometry, solver, frequency set or excitation count.

For the worked route, download and extract the complete inline microstrip example. Keep strip.ixi, process.ldf and strip.cir together in a new writable folder. The instructions below use PowerShell and the checked InductEx® 8.0.1/TTH 1.1.50 CLI. Confirm which executable your shell selects and its version before comparing results. The example process is a teaching stack, not a calibrated fabrication process.

Choose the setting to change

SegmentSize controls the maximum in-plane cuboid segment side length in the LDF Units. This example uses Units = 1e-6, so SegmentSize values 2, 1 and 0.5 mean micrometres. Cuboid layers can have their own SegmentSize override; inspect the full process before assuming the global value controls every layer.

Keep geometry, ports, material properties, frequency and solver settings fixed for this comparison. Keep the layer HFilaments values fixed too: those control height subdivision and are a separate refinement dimension. For triangle/tetrahedron meshing, SegmentSize applies at polygon edges and only its global value is used; the counts and costs below do not qualify those meshes.

AbsMin is not another name for maximum element size. The maintained manual defines it as a nonzero mapping radius around current-density nodes for cuboid meshing. Changing it can affect mapping, so leave it unchanged during this controlled SegmentSize sweep.

Create isolated cases

Run this preparation block from the folder containing the three extracted files. It creates three new case directories and changes only their copied SegmentSize line. Existing case directories cause an error so an earlier result is not silently reused.

$ErrorActionPreference = 'Stop'
$settings = @{'mesh-2um'='2.0'; 'mesh-1um'='1.0'; 'mesh-0p5um'='0.5'}
foreach ($caseName in $settings.Keys) {
    if (Test-Path -LiteralPath $caseName) { throw "Use a fresh directory: $caseName" }
    New-Item -ItemType Directory -Path $caseName | Out-Null
    foreach ($inputName in @('strip.ixi','process.ldf','strip.cir')) {
        Copy-Item -LiteralPath $inputName -Destination $caseName
    }
    $caseProcess = Join-Path $caseName 'process.ldf'
    $processText = [IO.File]::ReadAllText($caseProcess)
    $processText = $processText.Replace('SegmentSize = 2.0', 'SegmentSize = ' + $settings[$caseName])
    [IO.File]::WriteAllText($caseProcess, $processText, [Text.UTF8Encoding]::new($false))
}

Run and keep each result

1. Run each case with the commands below. Check a zero exit code and the current diagnostic output before accepting its numerical files. A leftover artifact or Job finished message alone is insufficient.
2. In each case, read output/elements.txt for the detailed L1 value and output/sol.txt for mesh information, time allocation, separately reported process peak memory, GMRES and SVD fit diagnostics. Keep these with the exact input files and executable/solver versions.
3. Inspect the generated mesh and compare geometry and electrical topology. If using the current Studio candidate, open the actual generated .msh; Live Mesh can retain an earlier Premesh preview (KB-GUI-009). The separate staged-run guide explains that inspection route.

Push-Location .\mesh-2um
inductex strip.ixi
$LASTEXITCODE
Pop-Location

Push-Location .\mesh-1um
inductex strip.ixi
$LASTEXITCODE
Pop-Location

Push-Location .\mesh-0p5um
inductex strip.ixi
$LASTEXITCODE
Pop-Location

Measured fixed-input comparison

These are actual values from one sequential Windows sweep. Time is the TTH-stage time reported by InductEx®, not total CLI elapsed time. Memory is the reported TTH process peak in MiB, not the combined whole-job or whole-system peak. On a narrow screen, scroll the table horizontally to read all columns.

Interpret the change before accepting convergence

Using abs(L_fine - L_coarse) / abs(L_fine), the 2 to 1 um step changed L1 by 0.61853%, and the 1 to 0.5 um step changed it by 0.54712%. The denominator is the finer result in both comparisons. Earlier single-refinement examples used the coarse denominator; retain the formula with any quoted percentage.

The second change is only modestly smaller. These three values do not establish a converged reference or a universal stopping threshold. Choose the required tolerance for the actual quantity and use further controlled refinement, including height subdivision where relevant, before accepting that requirement. If a change alters topology or produces an incomplete solve, investigate it before comparing values.

The 2 um result lies about 0.09028% below the unchanged 3.9 pH design target; the 0.5 um result lies about 1.24477% below it. Both are within 5% of the design value for this teaching case, but the coarser value being closer to the design does not make it more physically accurate. Design intent, process calibration and mesh convergence answer different questions.

All checked GMRES residuals were below their reported 1e-4 limit; the SVD fit relative errors were also small. These show the stated solver and lumped-fit behavior. They do not bound physical error or remove the observed mesh dependence. The SVD Unknowns = 1 belongs to the fitted one-element network; it is not the filament count of the field model.

Estimate cost on your own case

Use a small completed run on the intended machine as the first cost measurement, then refine in bounded steps. Record the actual filament count, solver-stage time, total elapsed time and available memory. Keep the geometry, frequency/excitation set, solver options and runtime identity with the record. Use the observed growth to decide whether the next run fits your resources; do not treat this strip's ratios as a capacity guarantee.

Here the finer 0.5 um case has about 14.5 times as many filaments as the 2 um case and about 12.2 times its reported TTH peak memory. The unchanged baseline repeat produced the exact same mesh bytes and L1, while its TTH time changed from 1.117 to 0.997 seconds and total CLI elapsed time from 4.770 to 1.795 seconds. Repetition matters when startup, licensing and host scheduling contribute to elapsed time.

Reported InductEx® and TTH peaks need not occur at the same instant; adding them is not a measured simultaneous peak. Leave room for the operating system, mesh preparation and other work. For an actual memory limit, measure the intended process set on the target machine and retain the measurement method.

If refinement fails or becomes too expensive

Keep the failed attempt separate and identify the earliest failed stage. Check the mesh aspect ratios, mappings, connected conductors and expected ports before changing residual limits. Preserve the last completed case as a comparison, then reduce the scope or return to its settings to check recovery. Do not accept a faster failed run as an optimization.

The checked baseline-repeat case recovered the exact original result and mesh after the sweep. It does not demonstrate recovery from an out-of-memory failure. For support, provide a shareable minimal model, exact version/settings, failing stage, count and cost records with private paths and licence details removed.