Research / Optimization

Four cables, one equilibrium

After Rafael Pastrana, Deniz Oktay, Kai-Uwe Bletzinger, Ryan P. Adams, Sigrid Adriaenssens,“Differentiable force density method for the design of lightweight structures”, Computer Methods in Applied Mechanics and Engineering (2026), doi:10.1016/j.cma.2026.118783.

Against jax_fdm (10 checks); 20 analytical/engine checks below
10 of 10 checks pass· see the checks
Against the paper’s figures
Our own inputs: compared with Eq. (7) and jax_fdm; the paper’s numerical results are not reproduced.

Research, not a design tool. Nothing here is an engineering result.

Find the shape of the four-bar system in Fig. 2 using positive force densities. Change the load, recompute in Rust, then open the found geometry as a truss in Stabileo.

Research content is available in English, Spanish and Portuguese. Other site languages display the English research content.

What is reproduced

This first version implements the forward equilibrium problem in Section 2.1, Fig. 2 and Eq. (7). Our four-cable net has one free joint and four fixed supports. Its topology follows the paper; its dimensions and loads are explicitly chosen here, not transcribed from the figure. The paper develops differentiable inverse design; that contribution is outside this version.

The forward method

Force density is axial tension divided by member length. Positive densities give an anchored network a positive-definite equilibrium matrix. The research crate assembles this matrix and uses the production engine’s sparse Cholesky factorization for the three coordinates. Forces follow from f = qℓ.

x* = (Σ qⱼ sⱼ + p) / Σ qⱼ

Supports are (±2, ±2, 0) m. The baseline downward load is 4 kN and every density is 1 kN/m. Eq. (7) then gives z = −1 m, length = 3 m and tension = 3 kN. These are derived targets for our inputs, not published numerical observations.

Evidence and tolerances

30 / 30 checks pass.

Five input cases compare every coordinate and cable force with unmodified jax_fdm in float64. Absolute tolerances are 1e-12 m for coordinates, 1e-12 kN for reference forces and equilibrium residuals, and 1e-9 kN for production truss forces. The validation file records each observed error. The replay is a load sweep computed by the same Rust solver, not convergence iterations. Scales are about 3 m, 3–12 kN and EA = 200000 kN. The 1e-12 budget allows float64 assembly and Cholesky roundoff; 1e-9 kN allows cancellation in recovering elastic forces from small displacements at this large stiffness.

Maximum reference position error: 2.2 × 10−16 m; Maximum reference force error: 0 kN.

Reference: jax_fdm, MIT, revision 9a17fd379071fb3511ca914f46e8d1e8c286460e. The manifest records the source archive SHA-256, output hashes and installed environment. This revision has not been confirmed as the publication revision.

What the production engine checks

The editor receives the found geometry, fixed support translations and the same vertical load as a 3D linear elastic truss. With equal member stiffnesses, symmetry makes its axial forces match FDM. Tests also check its vertical displacement against the elastic closed form. This is a separate analysis of initially unstressed bars: cable prestress and geometric stiffness are not transferred.

Not yet reproduced

No adjoints, gradients, optimization, shells, mixed tension/compression, shape-dependent loading, buckling or timing claims. No published optimized shapes or table values are reproduced. The preview is a labelled projection; the editor contains the full spatial model.

Research, not a design tool. Agreement with another code is not experimental validation or engineering certification.

Cite this

Stabileo Research, “Four cables, one equilibrium”, reproduction of Rafael Pastrana, Deniz Oktay, Kai-Uwe Bletzinger, Ryan P. Adams, Sigrid Adriaenssens (2026), doi:10.1016/j.cma.2026.118783. Implementation: research/form-finding/pastrana-2026 at commit fb241aa04936. Engine: the Stabileo engine at the same commit.