Skip to main content

How we know the numbers are right.

150 of the 254 methods are released on their own cited benchmark, 102 more are locked by regression tests, and 2 are held back. The finite-element solver is inside tolerance on 238 of 333 comparisons with vendor verification manuals, papers and closed-form solutions. Every comparison is public, the misses beside the passes.


The measured record.

Every figure carries its own denominator, and every row behind it is published.

Finite-element comparisons inside tolerance of an outside reference

238 of 333

Plus 42 self-checks. The other 53 rows are counted below, each with its reason

Published worked examples inside tolerance

44 of 46

The other 2 are outside tolerance and listed; each row names its source and the example or page

Methods released on their own cited benchmark

150 of 254

102 locked by regression tests, 2 held back

Automated engine tests, all passing at the last full run

9,253+

Invalid, boundary and worst-case inputs included

The 254 methods are the 233 analyses in the catalogue plus 21 interactive tools — the same analyses, not a different or larger set. A method is released when its own cited benchmark clears at least 20 assertions from at least 2 published sources. Expected values trace to the cited paper, standard or analytical solution. Open every row.


Build, prove, ship.

The order every engine is built in. Each step leaves evidence you can open: the source, the test, the benchmark or the limiting solution.

  1. Built from the source

    Every engine is coded from the governing standard, textbook or software theory manual, cited by clause. A chart factor with no closed form is an explicit input with a documented default, never a guessed curve-fit.

  2. Golden-tested on a hand-checked case

    Each engine ships with a hand-verified case plus the inputs that break weaker code: near zero, very large, water at the surface, a factor of safety of 1. It must hold at the criterion boundary.

  3. Benchmarked against published values

    Where a published worked example exists, its inputs are reproduced and its printed value asserted. Other checks pin the governing equation or a limiting solution. A failed benchmark is fixed in the engine, never by widening the tolerance.


Published cases we reproduce.

A slice of the benchmarks: published worked examples, closed-form and limiting solutions. Each row is held by an automated test or validation script against its cited source.

  • FE bearing capacity, undrained stripChecked against: Prandtl (1921)We reproduce: Collapse pressure (2 + π)·sᵤ = 5.14·sᵤ, within 0.4 % for smooth and rough footings
  • 2-D elasto-plastic footing (Hardening Soil, Mohr–Coulomb)Checked against: Prandtl (1921)We reproduce: Collapse load tends to 5.14·sᵤ (φ = 0) as the mesh is refined
  • 3-D elasto-plastic footingChecked against: Houlsby & Martin (2003) · Potts & Zdravković (2001) · Skempton (1951)We reproduce: Smooth circular footing: Nc, Nq and Nγ within 5 % of the published values; square φ = 0 shape factor 1.2
  • FE limit analysis (bounds)Checked against: Sloan (1988) · Prandtl (1921)We reproduce: Lower and upper bounds straddle 5.14·sᵤ
  • Transient unsaturated seepage (Richards)Checked against: Celia, Bouloutas & Zarba (1990) · van Genuchten (1980)We reproduce: Retention curve θ(ψ) of the Haverkamp sand, against the closed form
  • Wall surcharge, Boussinesq (line, point, strip)Checked against: NAVFAC DM-7.2, Fig. 11We reproduce: Published influence values σₕ·H/Q across m and n
  • Hardening Soil (triaxial and oedometer)Checked against: Schanz, Vermeer & Bonnier (1999)We reproduce: Berlin Sand III and Hostun sand parameter sets; the oedometer modulus law
  • Hardening Soil confined and layered settlementChecked against: Hardening Soil oedometer (closed form)We reproduce: Closed-form oedometer integral, within 2.5 %
  • 1-D consolidation (time rate)Checked against: Terzaghi (1943)We reproduce: t₉₀ = 0.848·H²/cᵥ (small-strain limit)
  • Finite-strain consolidationChecked against: Gibson et al. (1967) · CONDESWe reproduce: Power-law e–σ′–k benchmark case
  • 1-D site response (site period)Checked against: Kramer (1996), Ch. 7We reproduce: T₀ = 4H/Vₛ; transfer-function peaks at f₀, 3f₀ and 5f₀
  • Lateral pile p-yChecked against: Reese & Van Impe, Single Piles and Pile Groups under Lateral Loading · Matlock (1970)We reproduce: Published p-y deflection and moment cases
  • Vertical stress increaseChecked against: Boussinesq (1885)We reproduce: Closed-form influence factors
  • Liquefaction triggeringChecked against: Youd et al. (2001) · Idriss & Boulanger (2008) · four case-history databasesWe reproduce: Published CRR and factor-of-safety worked cases, and 867 case histories scored (below)
  • Layered-elastic pavement responseChecked against: Burmister (1945) · IRC:37-2018 (IITPAVE output)We reproduce: 30 printed subgrade strains (dual wheel) within 1 %
  • Rock strengthChecked against: Hoek & Brown (2002)We reproduce: Published GSI and σ₁ envelope cases

333 comparisons, every one listed.

Each row names its source (a vendor’s verification manual, a paper or a closed-form solution, with the page where there is one), the quantity, our value, the reference value where one exists, and the difference. Updated 14 November 2026.

Pass against an outside reference

238

Plus 42 self-checks against our own derivations

Outside tolerance

35

Kept in the registry with the reason

No verdict

14

6 with no published value to compare, 8 replaced by newer rows

Queued to run

4

Each can run with the solver as it is

Pass
Our value is within the tolerance the row states.
Outside tolerance
Our value is outside the stated tolerance, or the row records a known difference; it is not explained away.
Recorded, no published value
We ran the case, but the source publishes no number to compare with, so there is no verdict.
Superseded
An older stand-in or umbrella row, replaced by the newer row or rows named on it. It is not counted as a pass or as open.
Queued
The case can be run with the solver as it is, and has not been run yet.
Self-check
A pass against a value derived inside this project (our own hand calculation, a second solver of ours, or an identity). It is counted apart from the passes against an outside reference.

Every row that is not a pass carries its own reason on the verification page.

How this lines up with the options in other programs’ manuals, product by product, is on the comparison page.


867 sites where the ground did or did not liquefy.

The liquefaction triggering methods run on four published case-history databases, transcribed row by row with the page each row came from. No row is edited. Marginal cases are left out of the rates.

DatabaseCases transcribed
Idriss & Boulanger 2010 (SPT)230
Cetin et al. 2016 (SPT)210
Boulanger & Idriss 2014 (CPT)253
Moss et al. 2005 (CPT)174

Reading the SPT result. On the Idriss & Boulanger (2010) database, the Boulanger & Idriss (2014) method flags 110 of the 115 sites that liquefied (95.7 %) and clears 88 of the 112 that held (78.6 %). About one site in five that held is flagged: the method errs on the safe side. The report itself counts 7 liquefied sites below its curve and 18 non-liquefied sites above it, which is 93.9 % and 83.9 %. Plotting the transcribed rows against the same curve gives the same 7, and 23 in place of 18: 5 non-liquefied points sit within 5 % of the curve, where the table’s rounding decides the side. On the CPT database of Boulanger & Idriss (2014) the engine gives the report’s own counts exactly. The Cetin and Moss databases come with no printed success rate, so their rates are locked in regression tests and not quoted as published values.


Where we draw the line.

  • We never invent a benchmark number.Every value in a validation test carries its source: title, author or organisation, year. If we cannot cite it, it is not in the suite.
  • A failing benchmark is a bug, not a tolerance to widen.When a published case does not match, we fix the engine and re-run, or, if the source itself is wrong, we document why with evidence.
  • Frontier methods with sparse public numbers.For nonlinear finite-element work we gather every case we can rigorously cite and complete the set with exact analytical and limiting solutions, labelled as such. No figure is invented to fill a gap.
  • Standards are acquired legally.We implement methods and cite their clauses. Copyrighted code text, tables and figures stay with their publishers.

Measured scope, method by method.

Each method states where its measured range ends. A row leaves this list only with the test that closes it.

  • FE footing on frictional soilThe undrained strip reproduces Prandtl’s 5.14·sᵤ within 0.4 %, smooth and rough. On a φ = 30° strip on sand it misses Martin’s exact solution: 32.1 % low for a smooth footing and 5.9 % high for a rough one (the rough reading is on the unsafe side). The closed-form bearing methods run beside it as the check, and the modelling studio’s finite-element limit analysis brackets Martin’s value for both.
  • FE strength-reduction slopeACADS 1(a) at the default mesh: 0.961, inside the published cross-code envelope of 0.94 to 1.00 and 2 % below the limit-analysis reference 0.981. Bishop, Spencer and Morgenstern–Price land within 1 % of Slide2’s published values for the same case.
  • FE limit analysis (bounds)At the default refinement the bounds bracket 2+π between 4.07 and 6.25; the fine setting tightens them to 4.13 and 6.00. The modelling studio’s finite-element limit analysis brackets the Prandtl and Martin strip solutions in the registry.
  • Equivalent-linear site responseDriven by a target rock-outcrop PGA at the site frequency, for fast strain-compatible screening. Random-vibration site response runs from a Fourier amplitude spectrum.
  • Reduced PM4Sand-style elementA quick screening element in the catalogue. The full PM4Sand in the modelling studio is within 6 % of the authors’ published cyclic resistance at relative densities of 35, 55 and 75 %.
  • Coupled effective-stress dynamics (liquefaction)Scored against the VELACS Model 1 centrifuge record, the 2-D finite-element, particle-method and 1-D column results miss the pre-registered 15 % target: their pore-pressure error is 70.9 %, 29.1 % and 29.2 % of the recorded peak. We do not claim them as validated. Liquefaction triggering is scored on field case histories, above.
  • Finite-element verification registry35 of 333 rows sit outside tolerance. Each stays in the registry with its reason.
  • Unsaturated seepageThe catalogue’s transient Richards solver runs on a 1-D column and its 2-D seepage runs saturated. The modelling studio takes van Genuchten retention for each material in 2-D.

Run your own worked example.

Pick any analysis, enter the inputs from a case you already trust, and compare the number. The software should agree with the textbook; this page shows where it does and where it does not.