Quiet Gears: The Search Finds Helical
From a straight-cut brief to a 16.4° helix: seeded search re-derives the helical gain. The joint set cuts PPTE from 3.39 to 0.53 µm — both solids shown in 3D.
Overview
A shop needs a gearbox with a ratio and a quiet mesh. Input is four numbers: target ratio, design torque, max outside size, thrust limit. That brief stays fixed. The search starts at a straight-cut spur set and returns a full buildable set: tooth counts, module, shifts, tip trim, face width, helix angle, and finish grind.
Left shows the catalog pick at the same ratio. Right shows the grown set as true-scale 3D solids: straight sides versus helix twist. Use drag to turn them. There is no spin by design: micron relief would vanish under rigid turn.
Charts show TE over one mesh cycle, the tone spectrum with 5 tones, and the breed curve. Charts redraw each tick while the run moves.
Mesh Error
Mesh error is angular, not axial. If the pinion turns by , the gear should sit at exactly . Gap from that repeats each mesh cycle and pumps force at mesh tones. That force shakes shafts, case, and air. Less wobble means less whine.
A key point from prior work: grinding can raise mean error level while cutting wobble. Wobble is what counts here.
Why Helix Wins, and What It Costs
Three gains, all in the score: contact lines sweep slantwise instead of slamming flat; overlap adds to cross overlap so pairs share load; entry shock spreads in time. Code sums face slices as shifted spur contacts to get total stiffness. That math yields the smoothing from shape, not from a claim.
Cost rides along: axial push , checked against the brief limit. That trade is the choice the search must make.
Helix-Only Run
Phase A locks all at the catalog spur set and grows one gene: helix angle. One gene, one ask: does the search leave straight cut? It climbs to β = 16.4°, inside the car band, cutting worst fitness from 25.97 to 18.36 with thrust room to spare. Had it stayed at zero, this page would report that the score cannot see helical gain. It can see it. Thirty rounds trace the spur-to-helical path on a proxy, with limits stated.
Joint Set
Phase B grows all from the brief at once: shape, grind, and helix. Prior work finds joint growth beats step-by-step. The sample brief (ratio 1.5, 100 N·m, 100 mm size, 1500 N thrust) with 80 brief rounds returns 25/37 teeth, module 2.5, shift x1 = 0.30 with x2 = −0.107, 37.3 mm face, β = 11.2° with 0.92 overlap:
| Load | Catalog PPTE | Grown PPTE | Catalog H1 | Grown H1 | Push |
|---|---|---|---|---|---|
| 0.5× | 1.69 µm | 0.29 µm | 0.93 µm | 0.00 µm | 316 N |
| 1× | 3.39 µm | 0.53 µm | 1.87 µm | 0.00 µm | 631 N |
| 1.5× | 5.08 µm | 0.77 µm | 2.80 µm | 0.00 µm | 947 N |
Each load wins. No hidden trade. Worst fitness falls from 25.97 to 3.35. First tone reads 0.00 µm at all three loads in this proxy: slant contact kills the base tone. A small fourth tone near 0.24 µm stays. All numbers are proxy ranks with regression gates, not gauge reads.
Brief and Genes
Pressure angle stays 20°: stock hobs, swap parts, prior baseline. All else can grow: tooth counts snapped to ratio within 3%, module from {1.0–3.0}, shifts in [−0.3, +0.8], tip factor in [0.8, 1.25], face share in [0.3, 0.9], helix in [0, 35°], plus eight shape stations and shared crown. Hard caps price out bad sets: size, undercut, tip width above 0.25 module, contact ratio above 1.4, bend margin above zero, push inside limit. The bar line shows score, helix, push, and wobble.
Checks
Gates pin working angle at zero shift, undercut edge at 17 teeth, tip width and bend stress in hand bands, overlap and push in worked bands, helix tables matched to spur at β = 0, PPTE falling with helix, closed ring outlines, twist spans per gear, sealed solids, Phase A past 10°, joint run beating catalog by 20%+ with no load worse than catalog plus 15%, same-seed replay, step twins matched to full runs, plus a full in-widget design run with 3D fallback.
Honest Limits
Still proxy, rigid, quasi-static spur-and-helix. No friction. No dynamics. No gauge proof. It orders options. It does not call loudness in dB, and no dB number sits on this page for that reason. Slice math shows overlap gain, not true contact lines. Push is static math, not bearing life. All weights are tuned, not gauged. The 16.4° and the joint winner are best for the proxy, not for steel. A mesh-corrected core would move to wasm like the topology solver.
References
- Fonseca, Daniel J., S. Shishoo, Teik C. Lim, and D. S. Chen. 2005. “A Genetic Algorithm Approach to Minimize Transmission Error of Automotive Spur Gear Sets.” Applied Artificial Intelligence 19 (2): 137–151.
- Harris, S. L. 1958. “Dynamic Loads on the Teeth of Spur Gears.” Proceedings of the Institution of Mechanical Engineers 172: 87–100.
- Litvin, Faydor L., and Alfonso Fuentes. 2004. Gear Geometry and Applied Theory. 2nd ed. Cambridge University Press.