Special gears: non-circular, cycloidal, harmonic, nutating, and differential

A normal spur gear lives on a single promise: that the ratio between what goes in and what comes out stays constant, turn after turn, with no slip and no jam. The involute profile exists to keep that promise, which is why almost everything you design leans on it without a second thought. The gears in this article break that promise on purpose. Some change the ratio within a single turn; some drive it to numbers a spur train would need three stages to reach; some split one input into two outputs. In exchange they demand a geometry your printer doesn't forgive easily: conjugate profiles that have to match tooth for tooth, parts that flex thousands of times, eccentricities that throw the contact off center if tolerances stack up. It pays to understand what each one does before you ask the bed to print it in one piece.

diagram

The non-circular gear: the ratio changes within a single turn

Start with the easiest one to picture. In a pair of round gears the transmission ratio is the ratio of the pitch radii, and since those radii don't change, the ratio is constant. Make the wheels non-round — elliptical, or with a more arbitrary profile — and the pitch radius at the contact point stops being fixed: it grows and shrinks over the turn, and the ratio with it. The result is an output that cyclically speeds up and slows down even though the input turns at constant speed. That is exactly what you want when you need to generate a non-uniform motion law without resorting to a cam: a pump that delivers a pulsing flow, a mechanism that gives a fast stroke and a slow return, an indexer that comes to a gentle stop.

The condition you can't skip is that the two profiles be conjugate. At every instant the contact point has to lie on the line joining the two centers, and the sum of the two pitch radii at that point has to equal the center distance, always, at every angle. If the driving wheel is an ellipse turning on a focus, the driven one is not just any other ellipse: it is the curve that, at every position, completes that sum. Designing only one of the two wheels and "approximating" the other is a recipe for a mesh that seizes at some angles and rattles at others. The second wheel's profile is derived from the first and from the center distance; it isn't chosen.

Cycloidal: many contacts at once for backlash-free torque

The cycloidal reducer abandons the involute tooth entirely. In its place you have a disc with a profile of cycloidal lobes that rolls, pushed by an eccentric input shaft, against a ring of fixed pins. The disc doesn't turn about its own center: it orbits. The eccentric pushes it against the pins on one side, the disc rolls over them, and because the disc has one lobe fewer than the ring has pins, every full turn of the eccentric makes the disc walk back the equivalent of one tooth relative to the fixed ring. You pick up the output separately, through a set of output pins that pass through oversized holes in the disc and filter out the orbital component, leaving only the slow rotation.

What produces the reduction and what sets its magnitude are two different things, and it pays not to confuse them. The one-lobe difference is what produces the reduction: that is why the disc advances one step per turn and no more. But the magnitude of that reduction is set by the number of lobes on the disc. If the disc has N lobes, each turn of the eccentric rotates the output by 1/N of a turn: the ratio is 1

. The large reductions of a cycloidal — 1, 1 — don't come from the one-lobe difference, but from putting many lobes on the disc. The difference of one only decides that it advances one step at a time; how many steps fit in one turn is what sets the ratio.

The kinematic advantage is that many lobes are in contact with their pins at once. The load doesn't fall on one or two teeth as in a spur mesh, but is shared over a good part of the contour simultaneously. That gives the cycloidal its high torque in little volume and its very low backlash: with so many points in contact, there is no gap for the disc to swing freely through. The price is the eccentric orbital motion, which introduces a rotating mass imbalance whose force grows with the square of speed — that has to be balanced; that is why real cycloidals usually carry two discs at 180° to cancel each other. And even balanced, the load distribution over the pins produces a net radial force the eccentric's main bearing carries continuously.

In FDM, that load sharing over many contacts is exactly what suffers most from the printer's variability. If accumulated tolerances throw the disc off center relative to the pin ring, some lobes load up and others don't touch, and you lose at once the torque and the smoothness that justified fitting a cycloidal. On top of that, the eccentric bore and its bearings tend to be points of play that add backlash of their own. Here clearance calibration isn't a fine finish: it decides whether the mechanism works as intended or just rattles as a noisy eccentric. This ties directly into what Tolerances for moving parts explains: think in terms of clearance per side and measure your machine before you trust a table.

Harmonic (strain wave): a flexing ring that gets a big reduction in one stage

The harmonic, or strain-wave, gear achieves in a single stage what a conventional train would need several stages for: reductions of 100

and above, practically without backlash. It does it with three parts and an unusual idea: one of them works by flexing, not by turning rigidly. You have a flexspline, a thin-walled, elastic toothed ring; an outer rigid ring, the circular spline, with two more teeth than the flexspline; and inside the flexspline, an elliptical wave generator that deforms it, pushing its teeth against the rigid ring only at the two ends of the ellipse.

The kinematics are elegant and worth following slowly. The generator turns and walks the mesh zone around the perimeter. Because the flexspline has two fewer teeth than the rigid ring, each full turn of the generator makes the flexspline walk back exactly two teeth relative to the ring. That is where the reduction formula comes from: the ratio is the flexspline's tooth count divided by two. A 200-tooth flexspline gives 100

. Two teeth of walk-back per input turn — that's the reduction, and since the mesh happens by continuous deformation with many teeth in contact at each end of the ellipse, there is no appreciable backlash. You take the output from the flexspline; the rigid ring stays fixed (or the other way around, depending on the build).

The challenge in printing is the flexspline. That ring has to flex into an elliptical shape and back, twice per turn of the generator, for its whole service life. It is a pure fatigue part, and fatigue is the Achilles' heel of FDM.

The flexspline's wall thickness is the other critical number, but the figure that really rules isn't the absolute thickness but its ratio to the diameter. The maximum fiber strain when flexing into the ellipse is, to a good approximation, proportional to the thickness/diameter ratio: a 2 mm wall is very different on a 40 mm flexspline than on a 120 mm one. Watch that ratio, not the bare millimeter figure. And it pulls in two opposite directions: with the wall too thick for its diameter, you can't reach the elliptical shape without exceeding the material's allowable strain, and it cracks; too thin, it meshes poorly and buckles. The margin is narrow, and the material either narrows that margin or widens it. Nylon (PA) is the reference for fatigue under repeated flexing, and it's what you'll want in a serious flexspline. PLA fatigues early under cyclic strain and is a non-starter for long service life. PETG is tougher than PLA against impact, but its cyclic flexural fatigue is mediocre and it tends to creep under sustained load — exactly the case of a part that flexes twice per turn for life — so don't treat it as equivalent to nylon.

Nutating and differential: wobble and split

Two families remain, which I group together because they share the oddity, not the kinematics. The nutating gear combines rotation with nutation: the meshing element doesn't turn flat, but wobbles like a coin spinning to rest on a table, its axis tracing a cone. By tilting the toothed plate and making it nutate against a fixed ring, its mesh sweeps the circumference little by little. The reduction comes from the same idea as in the harmonic and the cycloidal: from the tooth count divided by the tooth difference between the nutating plate and the fixed ring. A small difference gives a large reduction in very little axial space, at the cost of a wobble motion that has to be contained and a contact that constantly changes zone. It's a geometry that's hard to model and harder still to print with clean contact, because the mesh plane is tilted and moving.

The differential is another animal entirely: it isn't after reduction, it's after splitting. It's the mechanism in a car's rear axle. One input — the ring gear — drives a carrier holding a set of spider gears, and those spiders mesh in turn with two output side gears. The point is that the input is split between the two outputs, letting them turn at different speeds: in a corner, the outer wheel travels farther than the inner one, and the differential allows that difference while still transmitting torque to both. The kinematic rule is clean: the average of the two outputs' speeds is set by the input. Brake one output and the other speeds up; if both turn alike, the assembly moves as a rigid block. That property — two outputs that compensate around an average imposed by the input — is what makes it useful far beyond the car: mechanical adders, motion splitting, controls that combine two rotations.

What they share on the print bed

Different as they are, two of these families share a clear print orientation, and it's worth knowing where it applies and where it doesn't. The non-circular gear and the cycloidal disc are flat parts: print them on the bed, with the gear's axis perpendicular to it. That is the orientation that gives the most profile fidelity, because the tooth contour is traced within the layer plane instead of stepping along the axis. A profile stepped by layers stops being conjugate, and a profile that doesn't conjugate seizes at some angles and gives a wrong ratio at others.

The rule, however, isn't universal, and it's important not to extend it blindly. The harmonic's flexspline is a deep cup, not a flat wheel; the nutating plate meshes tilted; the bevels of a classic differential have no flat orientation that preserves their profile. Each one asks for its own — the flexspline, vertical axis and a tough material; the differential, spur teeth so you can print it flat — and forcing the flat-wheel rule onto the parts that aren't flat is asking the bed for something it can't give you.

The second requirement is common, and in these mechanisms it weighs more than in any ordinary spur train: calibrated clearance. The cycloidal and the harmonic stake their torque and their absence of backlash on the load being shared over many simultaneous contacts; the non-circular stakes its mesh on the sum of pitch radii matching the center distance. In all three cases, an accumulated tolerance that throws the assembly off center breaks that sharing: some contacts load, others don't touch, and the mechanism loses exactly the property that made it worth the trouble. That's why here it isn't enough to pick a clearance from a table; you have to measure your machine's, as Tolerances for moving parts stresses.

The thread running through this whole family is that plastic flexes, fatigues, and drifts off dimension, and these gears force all three at once. The harmonic pushes fatigue to the limit, the cycloidal pushes accumulated dimension to the limit, the non-circular pushes conjugacy to the limit. When one of them fails on you, it will almost never be a "gear" failure: it'll be a wall that delaminated from being printed on edge, or a fastener that split a thin part at its weak point. That boundary between the clearance that lets it turn and the interference that cracks it is the same one that governs any tight fit, and it's developed in Interference without cracking.