A few months ago, I published my first article about QF Solver 0.2.7. At that time, the project was still quite early in its development.
Since then, I've spend quite a lot of time working on the solver. Some things were added, other parts needed corrections, and the verification process took more work than I initially expected.
On November 8, I released QF Solver 0.2.12 on PyPI.
I wanted to share what changed, what is actually supported by numerical evidence, and what still needs more work.
QF Solver is an open source finite element solver written in Python, mainly focused on structural mechanics. I'm developing it because I want to understand and inspect the equations, implementation and results rather than just running a model and looking at a stress plot.
I'm not claiming it can replace a mature industrial solver. There is still quite a lot to do before that would be a reasonable comparison.
An existing TET10 deformation example from the QF Solver repository.
More work on structural dynamics
One of the main areas I worked on is structural dynamics.
The starting point is the usual semi-discrete equation of motion:
M a(t) + C v(t) + K u(t) = f(t)
QF Solver already had dynamic examples, but I wanted to make some of the workflows easier to reuse and easier to check. Version 0.2.12 includes work on prepared systems, modal bases, reduced harmonic response and modal Newmark integration.
In modal reduction, the displacement can be approximated as u(t) ≈ Φ q(t), where Φ contains the selected modes. This reduces the dimension of the problem, but it also introduces an approximation that needs to be checked. The selected basis, damping model, loading and neglected modes can all affect the result.
The reduced dynamic routes are still experimental. A small residual in a reduced system isn't enough to prove that the full structural response is accurate.
An archived harmonic response example. It illustrates the frequency-domain workflow, not a new validation of every reduced harmonic method.
Newmark integration and time histories
I also continued working on transient response. Newmark integration is one of the available methods, including the average acceleration scheme with β = 1/4 and γ = 1/2.
For a suitable undamped linear system, it is possible to compare the numerical history with an analytical response and monitor energy drift and the dynamic residual. Those checks are useful, but their meaning depends on the exact configuration of the test.
Version 0.2.12 also includes work on modal Newmark, checkpoints and uniform base acceleration in specific routes. This does not mean arbitrary support motion or every transient configuration has been validated.
An archived direct Newmark example from the repository. It should not be confused with a complete validation of the newer modal route.
Some progress on nonlinear mechanics
Another area is nonlinear structural mechanics, especially selected J2 plasticity and geometrically nonlinear formulations.
I've been working on internal force calculations, tangent checks and the difference between trial and accepted states. This last part is quite important. If a nonlinear increment fails, a retry should not silently keep the material history of an increment that was never accepted.
It may sound like a small implementation detail, but it matters for consistent results and for debugging difficult cases.
There is still work to do here. Some of these formulations are only supported for specific configurations and the existing checks don't qualify every possible nonlinear coupling.
An existing J2 response example. It is not a general validation of finite strain plasticity.
Verification is probably where I spent the most time
I don't want QF Solver to become a collection of features that only work on a handful of demonstration models.
I'm trying to make the verification work more traceable, with documented inputs, numerical tolerances, reference quantities, residual checks and archived results. There are also cases compared against analytical solutions or other finite element implementations.
Something I've learned is that passing tests and having a validated numerical method are not the same thing. A software test can check that an API behaves correctly without proving the mechanical model is correct.
This is also why I'm keeping failed and inconclusive verification cases visible. I'd rather show a limitation than make the project look more mature than it actually is.
An archived solid element convergence example. Its interpretation depends on the chosen model and reference values.
The first QF Solver technical report on Zenodo
On November 14, I published a technical report on Zenodo for QF Solver 0.2.12.
It's called QF Solver 0.2.12: Numerical Verification Report, Discrete Spectral Comparison of a Weak-Bending Beam Model.
The report studies one discrete modal problem using three beam meshes with 115, 230 and 460 elements, six selected weak-bending branches, and reference and added-mass configurations.
The idea was to compare archived QF Solver modal results with a separately implemented reconstruction of the same discrete generalized eigenvalue problem. It is not a comparison to experimental measurements or proof of convergence to the continuous beam solution.
The paper includes a diagram of the model, frequency plots, error bounds, residual diagnostics and modal shapes. You can look at the original figures directly in the full PDF on Zenodo.
Figure 1, page 5: beam model and mesh configuration
For the scalar quantities, the results are encouraging. The report evaluates 36 frequency error bounds and 18 added-mass frequency-shift error bounds. All of them meet the stated scalar thresholds, provided the numerical spectral enclosures are reliable.
Figures 4 and 5, pages 11–12: frequency and shift error bounds
But there is an important limitation. The auxiliary eigenvector residuals do not satisfy their admissibility threshold, and a complete accepted correspondence of the compared modes has not been established for all meshes.
Figure 6, page 13: archived QF residuals versus auxiliary vector residuals
Because of that, the complete three-mesh verdict remains inconclusive. The scalar results should not be described as a full modal validation. The study also doesn't establish experimental validation or qualify the entire solver.
I think it's still worth publishing these results, including the parts that didn't meet the criteria. It gives a much more useful record of what was tested and what needs improvement.
The report is a technical preprint and has not been peer reviewed. The supplementary archive supports reproducibility of the postprocessing from archived data. It does not demonstrate a full rerun of the FEM assembly and eigenvalue calculations.
Installing version 0.2.12
The published package is available on PyPI:
python -m pip install "qf-solver==0.2.12" qf-solver --version There are examples and documentation on the project website and in the GitHub repository. Before using a calculation route, I recommend checking its documented scope and limitations. An example running without errors doesn't mean it is validated for every engineering use.
What's next?
I'm now working toward version 0.3.0.
The focus is on improving the numerical formulations, convergence studies, verification coverage and general code quality. I also want the documentation to make it easier for people to reproduce a result and understand where its assumptions come from.
QF Solver is still a small project, but it has changed quite a lot since I first wrote about it.
If you work on FEM, computational mechanics or numerical verification, feel free to take a look. I'm especially interested in feedback about difficult numerical cases or good independent benchmarks to investigate.




