In November 2026 an AI agent, Claude Code, wrote Python MCP client scripts that drove NovaSolver Studio, an open-source FEM/CFD app. It ran 187 solves and made 136 accepted comparisons: 132 with external references and 4 with another mesh of the same Studio model. The median difference was 0.13%.
Mesh-test or reference-check outcomes do not establish the cause of a gap. The gate requires converged levels, GCI at most 5%, sufficient element-count growth and a settled sequence; Studio's p >= 1 asymptotic flag is one route, not a general definition or a requirement for passing. Structural GCI generally tracks maximum displacement, rather than the local stresses or twist compared here; modelling, discretisation and reference assumptions may all contribute.
These are the ten most useful disagreements. Each comes with the number from the run and, where it fits, a line of Python.
1. Stokes drag is 6.5 times too low at Re = 200
Stokes' law, Cd = 24/Re, is exact as Re goes to 0. At Re = 200 it gives 0.12. Studio's flow solve gave 0.780; the standard drag curve (Clift, Grace and Weber) gives 0.776, a 0.6% difference.
Re = 200
print(24 / Re, 24 / Re * (1 + 0.1935 * Re**0.6305)) # 0.12 vs 0.7756 (Clift-Grace-Weber, 20 < Re < 260)
Even at Re = 10, Studio is 83% above Stokes. The inertia that Stokes neglects is already large there.
2. Saint-Venant's open-section formula over-predicts twist about 5x
Cut a slit along a tube and it becomes an open section with very little torsional stiffness. The Saint-Venant formula for an open thin-walled section assumes the ends are free to warp. With one end fixed, warping is restrained there.
For a 100 mm slit tube, the formula gave 0.491° of twist. Studio gave 0.095°. That is about a factor of 5. At 200 mm Studio's twist is 54% below the formula, at 400 mm 27% below: the longer the tube, the less the restrained end matters.
3. Thin-wall pR/t is 40% low when the wall is as thick as the radius
The hoop stress in a pressurised tube is often taken as pR/t. With R the inner radius and t = R, it gives 10.0 MPa where Lamé's thick-wall solution and Studio both give 16.67 MPa. It is low, which is the unsafe side.
ri, ro = 10, 20 # mm, t/R = 1
print((ro**2 + ri**2) / (ro**2 - ri**2), ri / (ro - ri)) # Lame 1.667 p vs pR/t 1.0 p
The full study (six wall thicknesses) found the choice of radius matters as much as the thickness: inner-radius pR/t is 5% low at t/R = 0.1 and 23% low at 0.5.
4. The straight-beam formula is 22% low on the inside of a C-ring
A C-shaped ring with R/h = 1.5. The straight-beam formula, N/A + M/Z, reads 22% lower than the Winkler–Bach curved-beam formula at the inner edge. Studio's FEM is 2.0% above Winkler–Bach.
Curvature shifts the neutral axis toward the centre of curvature and piles stress onto the inner fibre. The straight-beam formula does not know that.
5. 64/Re with a hydraulic diameter is 12% high in a square duct and 22% low in a 1:8 one
For laminar flow in a round pipe, f·Re = 64. Using the hydraulic diameter for other shapes is a common shortcut. It is not exact:
| Duct | Studio f·Re | Shah & London series | 64 |
|---|---|---|---|
| Square | 56.79 | 56.91 | 12% high |
| 1 : 8 rectangle | 81.44 | 82.34 | 22% low |
| Round | 64.37 | 64 | (exact) |
fRe = lambda a: 96 * (1 - 1.3553*a + 1.9467*a**2 - 1.7012*a**3 + 0.9564*a**4 - 0.2537*a**5)
print(fRe(1.0), fRe(1/8)) # Shah & London fit: 56.92, 82.36
The hydraulic diameter gets the scale right and the shape wrong, in both directions.
6. A 1 mm radius hole doubles the stress in a spinning disk
The textbook stress at the centre of a solid rotating disk is (3 + ν)/8 · ρω²b². Drill a hole of any size and the hoop stress at its edge jumps to twice that, in the limit of a tiny hole.
Studio: 1.996 MPa at the centre of the solid disk, 3.986 MPa sampled within 0.001 mm of the edge of a 1 mm radius hole. A 15 mm radius hole, 225 times the area, gave 4.038 MPa: only 1.3% more than the 1 mm one.
nu = 0.3
print((3 + nu) / 4 / ((3 + nu) / 8)) # hole edge (a -> 0) / solid centre = 2.0
The solid-disk formula is exact, for a disk that does not exist once there is a shaft hole.
7. van der Pauw's formula is 81% off on a sheet with a hole
The van der Pauw method measures sheet resistance with four contacts on the edge of a sample of any shape. Studio solved a 1.0000 Ω sheet as a square, a rectangle and an L-shape and the formula returned about 1 Ω, within 0.04% in every case.
With a hole in the sheet, it returned 1.8138 Ω. The theorem assumes a uniform material and thickness, point contacts and no holes. Shape does not matter; topology does.
8. The Euler buckling load can be 7x the load that yields the column
A 50 mm steel column, fixed at one end. Studio's linear buckling load was 164.8 kN. The load that makes it yield in plain compression is 23.5 kN. The elastic buckling load exceeds the nominal full-section yield load, so it cannot be used as a load capacity under the elastic assumption.
Studio's own Euler check flagged this case: the buckling load was 2.26% below the Euler formula, outside its 2% tolerance. Against Euler with shear deformation included, the difference is +0.25%. For the accepted columns at 100 mm and longer, the plain Euler formula agrees within 0.33%; the 800 mm case is excluded because its mesh test failed.
9. One Kt does not describe a thick plate
A plate with a central hole, hole diameter equal to the plate thickness (d/t = 1). The 2D stress concentration factor is one number. In 3D, the factor at the hole edge was 2.84 at mid-thickness and 2.46 at the surfaces: a 15% difference through the thickness.
Against Heywood's formula, Studio's net-section Kt was within 0.33% at d/W = 0.1 and 3.9% above at d/W = 0.6. Against Peterson's chart at d/W = 0.6 it was 1.9% above. A possible contributor is the 2D-versus-3D comparison, and at d/W = 0.6 the two textbook formulas themselves differ by about 2%.
10. Vlasov's thin-wall theory is 8% off when the wall is not thin
Back to the slit tube. Vlasov's theory includes warping restraint, which fixes most of item 2. But it assumes a thin wall, and here the wall is about 20% of the radius. Studio's twist was 8.3% above Vlasov at 100 mm, 5.1% at 200 mm and 4.0% at 400 mm.
Vlasov is closer here; its thin-wall assumption is a possible contributor to the remaining gap, whose cause was not isolated.
The pattern
These comparisons highlight assumptions that may contribute to the gaps; they do not rule out solver or modelling errors:
- Stokes: inertia neglected
- Saint-Venant, solid disk, van der Pauw: a boundary or topology that is not there
- pR/t, straight beam, Vlasov: a thickness that is assumed small
- 64/Re with D_h: a shape reduced to one length
- Euler: no yield
- Kt: no thickness direction
The agent found these by doing the boring thing at scale: solve, gate the mesh, compare with the formula, then ask why the gap is not zero. Formula assumptions are possible contributors; the checks do not isolate them from modelling or discretisation effects.
NovaSolver Studio is free and open source (GPL-3.0). It solves structural, thermal, flow and electrical problems in one app and can be driven from MCP clients.
- Overview and download: https://novasolver.jp/en/studio/
- Source code: https://github.com/novasolver/novasolver-studio
Which formula do you still use every week, knowing it is one step outside its assumptions, because it is close enough for the decision you are making? Tell us the formula and the case, and we may run it.


Top comments (0)