Columns that learned to dance: tree supports for robotic metal deposition

Dropped straight down 32 red columns land on the collar or the flange. The tree takes all 68 to the plate

Trees are not just an experiment. Cura, PrusaSlicer, Bambu Studio and Meshmixer ship them for plastic [1, 2, 3, 4], and Materialise added them to Magics for metal applications back in 2014 [5]. That is a good sign they work. What interests me is where they belong in robotic metal deposition, next to the other ways of holding an overhang, and how my generator grows them.

Left: a laser-cut rib fixture holds the part from the outside. Right: a deposited tree grows under the overhang, right where the torch needs a base

Deposited supports are built by the same program, in the same coordinates, and grow with the part. They reach recesses, take heat away and hold the part against distortion [7]. The price is arc time, wire and a saw.

They come in three shapes:

Left to right: columns, a wall and a tree on the same half-section of the demo part. Red marks where a support lands on the part

Columns are simple and predictable, and useless when anything of the part sits below. A wall is stiff, and in arc deposition it can be one continuous bead, which is a real advantage; the price is material and a big contact area to clean up. A tree touches the overhang with small tips, shares trunks and walks around obstacles; the price is slender branches and more maths.

Inside a pocket, the tilted torch hits the part

That freedom has a price. A positioner has to carry and precisely orient a heavy part, and inside enclosing geometry a tilted torch simply hits what was deposited before [8]. Collision checking and machine kinematics for multi-axis paths are hard [10, 11]. So many cells stay with planar layers and a fixed torch, which is simple and stable but safe only to about 15° of overhang [12], and robotic WAAM studies note that with flat deposition, near-horizontal overhangs may still need temporary supports [13].

My rule of thumb: tilt what you can, cut a fixture for what you can reach, and grow trees for the rest. That means overhangs above other parts of the model, pockets, and heavy parts you can't flip.

1. Find the overhangs. 2. Place the tips. 3. Merge on the way down. 4. Walk around obstacles. 5. Fuse everything into one closed mesh

1. Find the overhangs. A face needs support if it points down and tilts more than the surface angle from vertical. Faces lying on the plate don't, and very thin slivers along CAD face boundaries are ignored, because their normals can't be trusted. CAD cracks are welded shut first; open IGES sheets and scans are treated as two-sided surfaces, so they get supports too.

Faces are not the whole story, though. Tilt a flange a little, and every face around its lower rim is steep enough, yet the rim itself hangs in the air: the first layer of the flange starts there, with nothing below. The same goes for the tip of a corner pointing down. So the generator also looks for convex edges that point down and run flatter than the surface angle, and for lowest points, and puts contacts along them. Where such a rim just continues an overhang that already has tips, it reuses them.

2. Place the tips. Contacts are sampled on a grid with the chosen spacing, each a small gap below its surface. Two contacts count as duplicates only if they lie on the same local patch of surface, so two stacked surfaces a few millimetres apart each keep their own tips.

3. Merge on the way down. All tips form one front that descends layer by layer. On each layer, nodes that already carry more tips go first, and each tries its nearest neighbour first. Two nodes merge one layer lower at their weighted centre, weighted by how many tips each carries, if both can reach that point within the branch angle without touching the part. If no partner works, the node goes straight down, unless the part is in the way.

4. Walk around obstacles. When the way down is blocked, a beam search takes over (more below).

5. Fuse and check. Every branch becomes a tapered capsule, and all of them are fused into one closed mesh. Tips that can't reach the plate are listed, not silently dropped.

How the generator grows a tree

Two angles drive all of this, and they answer different questions. The surface angle decides where supports start. The maximum branch angle decides how far a branch may lean on its way down. Both come from the process, not from the algorithm: a fixed wire-arc torch is safe to about 15° of overhang [12], thin laser-DED walls stayed stable up to about 30° [14], and FDM is usually happy with 45° or more.

Left: faces tilted past the surface angle (amber) get supports; steeper ones (teal) don't. Right: every branch stays inside the cone set by the maximum branch angle

From every partial route, the search tries 25 places on the next layer: straight down, plus two rings of 12 directions whose radius comes from the branch angle. A route is expensive while the part is still below it; sideways distance and length add a little. The search keeps the 24 cheapest routes and, separately, the outermost route in each of the 12 directions. Without that, a route that has to crawl sideways under a broad ledge looks too expensive and dies, even when it's the only way out. Once the column below a route is clear, the rest of the drop is proven in one check. If nothing is found, the search runs again with half the step, and a trunk that got stuck after a merge is split back into its original tips, each with its own search.

Side view: the straight way down is blocked. Grey candidates are dropped, teal ones kept, and the outermost route (ringed) is kept on purpose until it clears the ledge. Top view: the 25 places tried from each route, every layer

A big flat underside, a heavy hot overhang, a tall slender branch and a sealed cavity

• Big flat undersides. They need many tips, and trees leave a worse underside there. Even consumer slicers switch to a normal grid under large flat overhangs [3]. In metal, a fixture, a wall or a plate is usually the better tool.
• Heavy or hot overhangs. Thin tips carry and conduct less. Supports also hold the part against distortion and take heat away [7], and less support volume can mean more distortion: in powder-bed fusion, tree-like supports show a clear trade-off between support volume and part displacement [18].
• Tall, slender branches. Sparse trees can be unstable and hard to print [19]. In wire-arc deposition a thin branch is a point-by-point bar, about 8 mm thick, built with pauses of 33 to 75 seconds per droplet in one study [20], and WAAM design rules warn that long, thin, unsupported members are difficult to deposit [21]. Bead width sets the thinnest branch; the process sets the steepest lean.
• No way out. A sealed cavity can't be supported from the plate by anything, tree or not. Branches only go down, so the generator marks such tips as unreachable instead of pretending.
• Six axes. If the part or the torch can be tilted, the best support may be none at all.

Left: 68 contacts and a forest of 256 branches. Tilted 59°: 17 contacts under the lower rims, 43 branches, and a part balancing on one edge

Before growing a forest, it's worth asking whether a rotation removes the overhang, so the extension also searches build orientations. It checks 84 near-uniform directions on a golden-angle sphere lattice [22], plus the part's big flat faces and the main axes.

On the demo part I asked for minimum supports. It tried 112 directions and found a 59° tilt. That took the part from 68 contacts to 17, from a forest of 256 branches to 43: 90% less support. The faces are now steep enough, so what's left holds the lower rims of the cap, the collar and the flange. The part loved it. It rose onto one edge of its flange, like a ballerina on pointe.

For arc deposition that is a terrible first layer: a tiny contact area, a poor path for the heat, and good luck clamping it. WAAM orientation research deliberately maximises first-layer area [23]. So the search gives a ranking, not a decision. The columns learned to dance; the part wanted to be a ballerina; the engineer still chooses the music.