Printing Through the Roll Gap: Deposition Paths Pre-Distorted for Canned Warm Rolling

When a powder is sealed inside a ductile metal container and rolled at a temperature below the melting point of its lowest-melting ingredient, it can reach near-full density without the furnace sintering that melts out volatile elements and leaves pores. This process is called canning rolling. It is attractive for alloys that ordinary sintering ruins. The cost is a very large, uneven plastic flow: an 80% thickness reduction stretches the compact several-fold along the rolling direction. Separately, additive manufacturing has shown that laying material along the directions in which a part will carry load, rather than in flat geometric rasters, can make parts many times stronger. These two facts collide. If you print a green part (the unconsolidated powder-plus-binder object) with load-aligned paths and then roll it, the roll gap rotates and bends those paths, so the alignment is lost.

We propose to plan the printed paths in the green state so that the rolling flow carries them into the desired load-aligned layout in the finished plate. In other words, we invert the consolidation flow instead of ignoring it. Most of that flow is a simple, nearly uniform stretch. The part the simple stretch gets wrong is concentrated in thin layers next to the container wall and at the plate edges. Local resolution enhancement, meaning spending fine computation only where the coarse answer is wrong, is therefore the natural way to make the inversion both accurate and cheap enough to sit inside a path optimizer.

Core Claim

For a printed, two-material green compact consolidated by canned warm rolling, the deposition path that produces a given final architecture is the pullback of that architecture through the rolling flow map $\phi$, which sends green coordinates $x$ to final coordinates $\phi(x)$. In the plate interior, $\phi$ is close to a plane-strain stretch. A path at in-plane angle $\theta$ to the rolling direction then obeys $\tan\theta_{\text{final}} = \tan\theta_{\text{green}}/\lambda$, where $\lambda$ is the longitudinal stretch set by the thickness reduction and the green-to-final density ratio. This follows from volume balance with negligible lateral spread. With $\lambda\approx 5$, a path meant to end at 45° must be printed at roughly 79°. Ignoring this is not a small error.

We claim three things. First, the residual error of this affine pullback is confined to boundary layers at the can–powder interface and the free edges. Second, a coarse global flow solve combined with refinement in those patches recovers the path layout that a uniformly fine solve would give, at a fraction of the cost. Third, plates printed with flow-inverted paths keep the stiffness advantage of stress-aligned deposition after full densification, while plates printed with directly aligned paths lose it.

This differs from every two-way bridge in the file. Stress-aligned planners such as the stress-flow non-planar method (Stress Flow Guided Non-Planar Print Trajectory Optimization), the swarm planner (Swarm-Based Trajectory Generation) and the adjoint fiber optimizer (More Stiffness with Less Fiber) all assume the printed geometry is the final geometry. The sintering surrogate (Virtual Foundry Graphnet) predicts deformation of isotropic green parts but is never inverted to design paths, and it covers unconfined furnace sintering rather than confined rolling. The 17-4PH extrusion study (Microstructural and mechanical properties of 17-4PH) links raster orientation to sintered strength, but only through flat-versus-on-edge heuristics.

Research Program

The material system comes from the one consolidation result that makes the premise viable. Canning rolling of Cu-50In-13Ga inside a copper can at 140°C reached 8.039 g/cm³, 99.7% of theoretical density. Sintering at 500°C instead exuded indium and left a porous compact near 7.8 g/cm³, which subsequent rolling made worse (Sintering and Rolling Behavior of Cu-50In-13Ga). That paper motivates the process and also constrains the design: any pipeline that sinters hot before rolling is ruled out.

We will extrude two powder pastes. The matrix is Cu-In-Ga. The reinforcing track is a Cu-rich, stiffer composition. Both are solvent-debound at low temperature, following the washing route of the 17-4PH study, and then canned and warm-rolled. Two materials are necessary rather than decorative. A single-material track architecture may simply vanish at 99.7% density, while a compositional contrast survives, can be measured by computed tomography or chemical mapping, and carries a stiffness difference.

The computational pipeline has three pieces. The first is a forward model of the rolling flow: a plane-strain-dominated plasticity solve of the can, compact and rolls. It runs globally on a coarse mesh. A refinement stage then re-solves small patches seeded where the coarse solution's velocity gradient or the can-contact shear is steepest. This borrows the scan-then-zoom logic of boundary patch refinement (Look Closer To Segment Better), whose oracle study found that fixing pixels within three pixels of boundaries was worth 17.8 AP. It also borrows the node-relocation approach of G-Adapt (G-Adaptive mesh refinement; G-Adaptivity: optimised graph-based mesh relocation for finite element methods), which moves mesh nodes toward steep gradients inside a differentiable finite-element solver.

To guard against truncated boundary conditions, a known weakness of tight patch crops, each patch receives the coarse solution as its boundary data. The refiner predicts a residual on top of the coarse map, in the spirit of PatchRefiner (PatchRefiner). That paper's own limitation also constrains us: a refiner cannot recover a feature the coarse pass never saw. We therefore keep a cheap global error indicator that can trigger new patches.

The second piece is inverse design. We treat the green path node coordinates as design variables, as in the adjoint fiber optimizer. The objective is the stiffness of the final machined plate evaluated on $\phi$(paths). Gradients flow through the forward flow model, so the forward model is called hundreds of times. That repeated use is why it must be cheap.

The third piece is the target architecture. For a given service load on the final plate, we compute stress-aligned target paths with the swarm planner, then initialise the inverse problem from their affine pullback.

Baselines

The naive reference prints the target stress-aligned paths directly in the green state with no correction, then rolls. It shows what happens when prior planners are used unchanged in this process. The geometric-heuristic reference prints concentric or ±45° rectilinear paths. It anchors the mechanical comparison, since the adjoint fiber paper reports its optimized paths as roughly 1.6× stiffer than concentric rings on a two-hole plate.

The strongest simple competitor is the affine pullback, the closed-form angle correction above. It costs nothing and may capture most of the effect. The accuracy upper bound is a uniformly fine plasticity solve of the same rolling pass, used to pull back paths. All flow models share the same mesh family, material law, friction calibration and wall-clock budget per optimizer iteration. The refined model beats this baseline only if it matches the fine solve's final-plate path placement within print-line width while running at least an order of magnitude faster inside the optimization loop.

The strongest existing path-design method is the swarm planner applied in the final geometry. It reports about 10% higher specific modulus than naive load-aligned rectilinear infill on open-hole tensile specimens, and its 198 ms generation time makes it the realistic incumbent. We give it the same final load case and print its output both uncorrected and affinely pulled back.

The key ablation removes patch refinement, leaving coarse-only pullback. A second ablation spends the same number of refined cells at random locations rather than at the can boundary and edges. Together they isolate whether the gain comes from where the resolution goes or simply from having more of it.

There is no reported result on Lacuna for path-orientation fidelity after canned rolling, so the success thresholds for path placement are set relative to our own uniformly fine solve.

Why not just…?

Why not just print the target paths and let rolling do the aligning? Rolling rotates every off-axis path toward the rolling direction by the $\tan\theta/\lambda$ relation. Rolling therefore "aligns" everything with one direction, whatever the service load demands. For a uniaxial load along the rolling direction this might suffice, and we include that load case precisely to identify where flow inversion is unnecessary. For loads around a machined hole, the stress-flow and swarm results show that orientation matching the local stress field is what carries the benefit. The naive baseline measures how much of that benefit is lost.

Why not just use the affine pullback? This is the real threat. If the boundary-layer error is narrower than one printed track, the closed form is the method, and we will say so. The first experiment settles this directly. It measures, on sectioned rolled plates, the deviation between affinely predicted and realised track positions as a function of distance from the can wall and edges. Refinement earns its place only if that deviation exceeds track width over a region that matters mechanically, for example the material ending up next to a hole machined near an edge.

Why not just run a uniformly fine plasticity solve? For a single forward pullback, a fine solve is affordable and we use it as ground truth. The case for local refinement rests entirely on the inverse problem. The adjoint optimizer calls the forward model repeatedly, and if fine solves stay tractable at that call count, we drop the refiner. We report wall-clock to convergence for both, against G-Adapt's reported 21.01% error reduction in 88 ms versus about 12.7% for classical Monge–Ampère relocation on Poisson. That figure suggests, but does not establish, that learned relocation transfers to plasticity.

Why not just sinter conventionally and HIP afterward? For this alloy family the evidence is direct. Hot sintering exudes indium, and subsequent rolling dilates rather than closes the resulting pores. Even where HIP is used, crack-free results on MgB₂ required holding pressure through cooling (Hot isostatic pressing of bulk magnesium diboride). A plain HIP cycle is therefore not a safe default.

Proposed Experiments

The first experiment is a survival test. Plates of straight two-material tracks at 0°, 30°, 60° and 90° are printed, debound, canned and rolled at 140°C to 80% reduction. Density is compared against the 8.039 g/cm³ reference. Track angle and position are measured by tomography and chemical mapping and compared with the affine prediction. If tracks smear beyond recognition, the program stops here, and that outcome is itself worth reporting.

The second experiment characterises the boundary layer. It maps affine-pullback error versus distance from the can wall and edges, and checks whether coarse-plus-patch and uniformly fine solves reproduce it. The third is inverse design on the two benchmark geometries from the stress-aligned literature: the open-hole tension specimen of the swarm planner and the two-hole plate of the adjoint optimizer, with holes machined after rolling. Every baseline and ablation above is printed and tested for stiffness and failure load.

The decisive check has three parts. Flow-inverted plates must exceed both uncorrected swarm paths and concentric paths in specific stiffness after rolling. They must match fine-solve pullback within one track width near the holes. Finally, the edge-focused-versus-random refinement ablation must show that resolution placed at the can and edge boundary layers, not resolution in general, accounts for the gap over the affine pullback.