Optimizing 3D-Printed Fiber Paths for Strength, Not Just Stiffness

Computer-designed parts are usually optimized with a simulator that treats the material as perfectly springy: stress rises in proportion to strain forever, nothing yields, and nothing breaks. This assumption is called linear elasticity. It is a reasonable way to design for stiffness. It says almost nothing about when a part fails. A clear example is continuous carbon-fiber 3D printing. A recent pipeline plans the fiber routes by following gradients through a linear-elastic finite-element simulation, and its printed parts come out about 1.6 times stiffer than the commercial slicer's concentric rings. The same pipeline never predicts, and never measures, how much load those parts carry before the nylon matrix yields, the fiber pulls out of the matrix, or the fiber snaps. We propose to route fibers so that the part's measured failure load is as high as possible. The optimizer would work through a simulator that models matrix yielding and progressive fiber and interface damage. That simulator would be calibrated with uncertainty against coupon tests, and every design would be checked by breaking printed parts.

The central bet is that the stiffest fiber layout and the strongest one differ. Stiffness-optimal routes pack fiber along the main tension lines and pass close to holes and corners. Linear theory puts a stress spike near those features. In a ductile nylon matrix that spike spreads out once the material yields, so the actual failure sequence depends on behavior the stiffness optimizer cannot see. If the bet is right, stiffness-optimized parts will be clearly weaker than parts routed for failure load at equal fiber length. The test is physical: measured peak load and energy absorbed to failure on an Instron 600DX.

Core Claim and Prior Work

Two papers supply the method we build on and, through what they leave out, the motivation. More Stiffness with Less Fiber uses gradients computed by the adjoint method (a way to get the derivative of an objective with respect to every design variable from one extra linear solve). It pushes spline-parameterized fiber paths toward minimum 2D linear-elastic compliance, starting from a greedy walk along principal tension directions and refining at increasing resolution. Gradient-Based Dovetail Joint Shape Optimization for Stiffness applies the same adjoint-plus-physical-test approach to interlocking joints, with a smooth contact penalty. Both validate only initial stiffness, and both state that yielding, delamination, and rupture are out of scope. We keep their parameterization, multi-resolution schedule, printer, and test machine. We replace the objective and the physics.

Classical elasticity explains why the replacement matters. Solutions for perforated thin plates and press-fitted shafts produce stress concentrations that local yielding redistributes in real parts. An objective built on the linear peak stress will therefore misjudge where the part actually fails.

Another paper warns against the obvious implementation. DLO-Lab found that gradients through plasticity and contact become rugged, and that a sampling optimizer, the covariance matrix adaptation evolution strategy (CMA-ES), beat gradient methods overall. Its reported average success rate was 86.6%. We treat this as a real design constraint, and it shapes the hybrid optimizer below.

For calibration we borrow from Learning to Simulate Tree-Branch Dynamics for Manipulation. That work infers nonlinear stiffness and damping as a posterior distribution with prior-guided Stein variational gradient descent, a particle method for approximating that distribution. Its 95% intervals covered 93% of field trajectories. We carry this approach to plasticity and damage parameters. Unlike the segmental bridge pier study, which calibrated damaged plasticity deterministically from two specimens, we fit to repeated coupons and propagate the uncertainty into predicted failure loads.

Research Program

Simulator. The simulator is a 2D plane-stress finite-element model of the printed layer. The nylon matrix follows von Mises (J2) plasticity with isotropic hardening. Each fiber is an embedded bar joined to the matrix by a cohesive interface with a traction–separation law. Fiber rupture is represented as a smooth damage variable. Load is applied in small displacement increments up to and past the peak.

Gradients. Gradients come from the path-dependent adjoint. It is solved backward through the converged load increments, as in the dovetail and fiber papers but now with history variables. To keep this adjoint usable, softening is regularized with a length scale tied to the fiber width, which prevents mesh-dependent localization.

Objective. The objective is a smoothed peak load (a soft-max over load increments) or energy absorbed up to 20% load drop. The total fiber length is held fixed.

Optimizer. The optimizer follows the DLO-Lab lesson. At the coarsest spline resolution, a few dozen control points, CMA-ES explores globally using the nonlinear simulator. Adjoint descent then refines the result as the paths are upsampled to hundreds of control points. At that resolution sampling becomes too expensive.

Calibration. We will print unidirectional, ±45°, and open-hole coupons and load each to failure. From these tests we infer a posterior over matrix yield stress, hardening, interface strength and toughness, and fiber strength. The prior weakly ties interface toughness to matrix yield.

Uncertainty on candidate designs. Every candidate design is evaluated under posterior samples of the parameters. Each design therefore has a predicted failure-load interval before it is printed.

Baselines

The naive reference is the Markforged slicer's concentric-ring fiber layout at the same fiber length. It is what practitioners actually print, and it anchors the rest.

The strongest existing method is the linear-elastic compliance-optimized path of More Stiffness with Less Fiber. That paper reports its paths to be "roughly 1.6x stiffer" than concentric rings on a rectangle with two holes, and to match baseline stiffness with "70% less fiber" on SketchGraphs-derived silhouettes. We found no reported result on Lacuna for the failure load or energy to failure of those paths. Measuring them is itself part of the contribution. We will re-implement their pipeline with the same initialization, resolution schedule, and fiber budget as ours, and give it an equal compute budget in wall-clock time. Beating it means a higher mean measured peak load, with non-overlapping 95% bootstrap intervals over five specimens per design, on at least three of five geometries. It also means stiffness no worse than 15% below theirs, since a strength gain that gives up all stiffness would not be interesting.

Two further baselines isolate the key ingredient, which is post-yield physics. The first is a linear-elastic strength surrogate: the same adjoint pipeline minimizing a smoothed maximum (p-norm) of a Tsai–Wu-style failure index plus an interface-stress index. It represents the classical "linear analysis plus failure criterion" route. The second is a derivative-free search: CMA-ES alone over low-resolution control points, calling the same nonlinear simulator with the same budget. It tests whether gradients add anything at high resolution.

Why not just…?

Why not just use linear elasticity with a failure criterion or safety factor? This is the most serious threat, and we test it head-on with the strength-surrogate baseline. The mechanics literature above suggests that a linear stress peak at holes misstates where failure starts once the matrix yields. However, if the Onyx matrix is brittle enough in practice, the surrogate may match the nonlinear optimizer. If the two produce statistically indistinguishable measured failure loads, we will report that the extra physics is unnecessary for this material and focus the contribution on the calibrated strength measurements.

Why not just take the stiffness-optimal paths, which may already be strong? It is plausible that routing fiber along tension lines also delays rupture. Breaking the baseline's parts answers this directly. We will also check whether the simulator reproduces the observed failure locations, from digital image correlation and post-mortem photos, for both families of designs.

Why not just run CMA-ES or Bayesian optimization over commercial nonlinear finite-element software? For the few-parameter dovetail spaces this probably works as well as gradients, which is why we chose fiber routing, where designs have hundreds of degrees of freedom. The pure-CMA-ES ablation measures where sampling stops keeping up as resolution grows. If it matches the hybrid at every resolution under an equal budget, we will concede that the adjoint is unnecessary.

Proposed Experiments

The geometries are the rectangle with two holes plus four SketchGraphs silhouettes from SketchGraphs, matching the fiber-path paper's benchmarks, loaded in tension and three-point bending. Each geometry–method pair gets five printed specimens, loaded to failure on the Instron 600DX with digital image correlation. The metrics are peak load, energy to 20% load drop, initial stiffness, failure location, and the coverage of the predicted failure-load intervals.

There are three decisive checks. First, whether routing optimized for failure load beats compliance-optimal routing on measured peak load at equal fiber length. Second, whether the calibrated nonlinear simulator ranks all designs in the same order as the physical tests do (Spearman correlation above 0.8), when the linear surrogate does not. Third, whether the 95% predicted intervals cover close to 95% of measured failure loads on held-out geometries. If the first check fails, the idealization is harmless for this process. That is a useful negative result, and one no prior work in the file has measured.