Spiral bevel gears carry torque between intersecting shafts in helicopter main and tail rotor drives, where low mass, high tooth strength and quiet operation are demanded simultaneously. Their teeth are cut on generating machines whose settings determine a surface that is conjugate to the mating member at only one point; contact is therefore localized and the meshing behaviour away from the reference point is a consequence of the machine settings rather than a directly specified quantity. Under torque, the teeth deflect, the instantaneous ratio departs from its nominal value, and the resulting rotational error — the loaded transmission error (LTE) — becomes one of the dominant periodic excitations of the gearbox. Its peak-to-peak amplitude, not its mean level, governs the excitation strength: a curve that repeats with large swings every mesh cycle radiates noise at the mesh frequency and its harmonics.
Classical Gleason local-conjugate design fixes the contact pattern empirically. Litvin’s local synthesis gave the designer three second-order quantities at the reference point — the direction of the contact path, the first derivative of the transmission-ratio function, and the semi-axis of the instantaneous contact ellipse — from which the principal curvatures and directions of the pinion are derived and the machine settings back-calculated. Because the surface is smooth and continuous, these parameters are controlled only in the neighbourhood of the reference point; far from it the contact path can bend sharply and the contact ellipse can change length abruptly. Subsequent work extended the control to the toe, middle and heel of the tooth, removing third-order defects and giving global rather than local control of the mesh. All of these procedures remain indirect: settings are chosen, the resulting surface is evaluated, and the loop is repeated until the design targets are approximately met.
Ease-off flank modification reverses this logic. The target surface is prescribed first, from functional requirements — contact pattern position and extent, transmission error shape and amplitude — and the machine settings are then solved so that the manufactured surface reproduces the target. The gap between the target surface and the conjugate (or reference) surface, measured along the common normal, is the ease-off. Designing in this space makes the transmission error a directly assignable function instead of an outcome.
Why a Second-Order Parabolic Correction Stops Working Above Contact Ratio Two
Tooth modification aimed at reducing LTE works by compensating the deflection of the tooth pair at the positions where the mesh is stiff, so that the sum of geometric and elastic contributions stays flat. A parabolic (second-order) transmission error curve has produced good vibration results in gears with contact ratio below two. In that regime a single tooth pair passes through double contact, then single contact, then double contact again as it moves from engagement to disengagement. The mesh stiffness therefore has one broad minimum in the middle of the cycle, and a single concave parabola matches it.
When the contact ratio lies between two and three, a tooth pair traverses three-pair, two-pair, three-pair, two-pair and three-pair contact before leaving mesh. The number of simultaneously loaded pairs changes four times per cycle, so the mesh stiffness and the load sharing between pairs vary with a different, higher-frequency signature. A parabola provides only one region of curvature and cannot follow this pattern; a residual ripple survives whatever the amplitude of the parabolic correction. What the correction curve must reproduce is the sequence of stiffness transitions set by the contact ratio, which calls for a polynomial of substantially higher order whose slope changes at the same angular positions as the pair-count changes. One useful design guideline follows from this reasoning: for a given gear pair, the LTE amplitude passes through a local minimum somewhere in the interval 2 < ε < 3, so the design contact ratio should be searched inside that band rather than fixed a priori.
A Seventh-Order Concave Transmission Error Target
The target adopted here is a mid-concave (wave) curve, symmetric about the middle of the mesh cycle, written as a rigid-body rotation plus a seventh-order polynomial perturbation:
$$\begin{cases}\varphi_2=\varphi_2^{(0)}+\dfrac{Z_1}{Z_2}\left(\varphi_1-\varphi_1^{(0)}\right)+\delta\varphi_2\\[4pt]\delta\varphi_2=a_0+a_1\left(\varphi_1-\varphi_1^{(0)}\right)+\cdots+a_7\left(\varphi_1-\varphi_1^{(0)}\right)^{7}\end{cases}$$
Here $Z_1$ and $Z_2$ are the pinion and gear tooth numbers, $\varphi_1$ and $\varphi_2$ the actual rotation angles of pinion and gear, $\varphi_1^{(0)}$ and $\varphi_2^{(0)}$ their initial values, and $a_0,\dots,a_7$ are the polynomial coefficients that define the shape. The seventh degree supplies enough free parameters to place the slope reversals of the concave curve at the four pair-count transitions of a 2 < ε < 3 mesh while keeping the curve smooth and symmetric; $a_0$ shifts the whole curve and does not affect its peak-to-peak amplitude.
From Target Transmission Error to an Ease-Off Tooth Surface
The gear tooth surface is treated as an imaginary cutter. The pinion auxiliary surface is generated from it in the coordinate system of the cutting machine, with the rotation of the two members coupled through the prescribed geometric transmission error, so that gear and auxiliary pinion are in line contact and the geometric transmission error of the pair equals the prescribed function exactly:
$$\begin{cases}\mathbf{r}_1’=\mathbf{M}_{1h}(\varphi_1)\,\mathbf{M}_{h2}(\varphi_2)\,\mathbf{r}_2\\[3pt]\mathbf{n}_1’=\mathbf{L}_{1h}(\varphi_1)\,\mathbf{L}_{h2}(\varphi_2)\,\mathbf{n}_2\\[3pt]f_{12}\left(s_g,\theta_g,\phi_g,\varphi_1\right)=0\end{cases}$$
where $\mathbf{M}$ and $\mathbf{L}$ are the homogeneous transformation and rotation sub-matrices linking the gear, the machine cradle and the pinion, and $f_{12}=0$ is the equation of meshing with $s_g$, $\theta_g$ and $\phi_g$ the gear surface parameters. A straight contact path is then laid out on the auxiliary pinion surface to fix the design contact ratio:
$$y=\left(x-x_M\right)\tan\left(\gamma_1+\eta-\Gamma_1\right)+y_M$$
with $x_M,\,y_M$ the coordinates of the reference point $M$, $\eta$ the angle between the contact path and the root cone, $\Gamma_1$ the pinion pitch cone angle and $\gamma_1$ the pinion root cone angle. Finally the auxiliary surface is discretized into $k$ grid points and each point is displaced along its own unit normal by a correction $\zeta_i$, producing the ease-off target surface:
$$\mathbf{p}_i^{*}=\mathbf{p}_i^{0}+\zeta_i\,\mathbf{n}_i^{0},\qquad i=1,\dots,k$$
The distribution of $\zeta_i$ is what converts the prescribed concave transmission error into a physical gap field on the tooth; a positive value removes material, a negative value adds it, and the resulting normal gap varies smoothly from toe to heel and from root to tip.
Optimization Formulation
The pinion is the member whose surface is modified, so its machine settings form the design vector $\mathbf{d}$: radial cutter position, initial cradle angle, ratio of roll, vertical and horizontal offsets, bed position, machine tilt angle, and the second- through seventh-order modified-roll coefficients — thirteen variables in total. Two objectives are minimized simultaneously:
$$\begin{cases}\min f_1(\mathbf{d})=\max\left(T_e\right)-\min\left(T_e\right)\\[3pt]\min f_2(\mathbf{d})=\mathbf{h}(\mathbf{d})^{\mathsf{T}}\mathbf{h}(\mathbf{d})\end{cases}$$
where $T_e$ is the loaded transmission error and $h_i(\mathbf{d})=\left(\mathbf{p}_i^{*}-\mathbf{p}_i\right)\cdot\mathbf{n}_i$ is the signed normal deviation between the achieved pinion surface and the ease-off target at grid point $i$, assembled into the vector $\mathbf{h}$. The first objective suppresses the vibration excitation directly; the second keeps the manufactured surface faithful to the target so that the prescribed transmission error is actually realized. Two constraints apply. The design contact ratio is restricted to the interval [2, 3] for the reason given above. In addition, the geometric and loaded transmission error curves must intersect somewhere in the mesh cycle; if they do not, the tooth pair separates and re-enters contact at the tooth edges, producing edge contact and severe impact loading.
No closed-form relation exists between the machine settings and either objective — each evaluation requires a loaded tooth contact analysis — so the problem is solved with the non-dominated sorting genetic algorithm NSGA-II. Its fast non-dominated sorting, crowding-distance operator and elitism preserve diversity along the Pareto front while keeping the number of contact analyses manageable, which linear weighted-sum formulations cannot do without arbitrary weighting and a tendency to converge on local optima.
Case Study: a 27/74-Tooth Aerospace Gear Pair
The procedure was applied to the working flank of an aerospace spiral bevel gear pair: 27 teeth on the left-hand pinion, 74 teeth on the right-hand gear, face width 40 mm, shaft angle 87°, module 3.85 mm, evaluated at a torque of 950 N·m and an input speed of 8000 r/min. To make the comparison meaningful, a conventional second-order (parabolic) design was first optimized under exactly the same objectives and constraints, so that the wave design is compared against the best achievable parabolic design rather than against an arbitrary baseline.
| Parameter | Gear (initial) | Pinion (initial) | Pinion, 2nd-order optimized | Pinion, wave design |
|---|---|---|---|---|
| Radial cutter position / mm | 119.687 | 113.435 | 113.435 | 112.547 |
| Initial cradle angle / (°) | −43.568 | −43.276 | −43.276 | 43.098 |
| Ratio of roll | 1.083 | 2.841 | 2.841 | 2.820 |
| Vertical offset / mm | 0 | −5.210 | −5.210 | −6.024 |
| Horizontal offset / mm | 0 | −3.110 | −3.110 | −3.594 |
| Bed position / mm | −1.741 | 0.409 | 0.409 | 0.565 |
| Machine tilt angle / (°) | 66.005 | 18.897 | 18.897 | 18.897 |
| 2nd-order modified-roll coefficient | 0 | 0.044 | 0.044 | −0.006 |
| 3rd-order modified-roll coefficient | 0 | 0.113 | 0.113 | 0.037 |
| 4th-order modified-roll coefficient | 0 | 0 | 0 | 2.390 |
| 5th-order modified-roll coefficient | 0 | 0 | 0 | −5.651 |
| 6th-order modified-roll coefficient | 0 | 0 | 0 | −60.197 |
| 7th-order modified-roll coefficient | 0 | 0 | 0 | 280.570 |
The wave design activates all four additional modified-roll coefficients and shifts the offsets by fractions of a millimetre, which is the expected signature of a surface correction that has to follow a higher-order transmission error: the lower-order terms alone cannot reproduce the required inflection pattern. The deviation between the achieved pinion surface and the ease-off target remained small over the whole grid, confirming that the machine is capable of realizing the prescribed wave surface.
Results: Loaded Transmission Error and Dynamic Load Factor
For the parabolic design, optimization reduced the peak-to-peak LTE amplitude from 3.502 arcsec to 2.306 arcsec, a 34.152% reduction, while the geometric transmission error curve remained parabolic as intended. The wave-designed pinion, evaluated under identical conditions, produced an LTE amplitude of 0.888 arcsec — 61.492% below the optimized parabolic design and roughly 75% below the unoptimized baseline. In both cases the geometric and loaded transmission error curves intersected, and the computed contact patterns stayed inside the tooth boundaries, so no edge contact occurred at the rated load.
| Design | LTE amplitude / arcsec | Reduction relative to its own baseline | Dynamic load factor |
|---|---|---|---|
| 2nd-order, before optimization | 3.502 | — | 1.117 |
| 2nd-order, optimized | 2.306 | 34.152% | 1.081 |
| Wave design | 0.888 | 61.492% (vs. optimized 2nd-order) | 1.019 |
The dynamic load factor $K_v$, which aggregates the internal additional load generated by pinion and gear mesh vibration, follows the same trend: 1.117 before optimization, 1.081 after parabolic optimization, and 1.019 for the wave design. The factor was obtained from an eight-degree-of-freedom lumped-parameter model of the pair on elastic supports, with translational stiffness and damping $k_{1x},k_{1y},k_{1z}$, $k_{2x},k_{2y},k_{2z}$ and $c_{1x},\dots,c_{2z}$ along the three axes of each member, a time-varying mesh stiffness $k_n$, mesh damping $c_n$, and torques $T_1$, $T_2$ acting on the torsional coordinates $\theta_1$, $\theta_2$; friction and torsional–lateral coupling were neglected. A drop of the load factor from 1.117 to 1.019 corresponds to a substantially smaller internal dynamic increment, consistent with the halving of the LTE amplitude.
Finite Element Cross-Check
Because the loaded tooth contact analysis rests on a compliance model of the tooth pair, the wave design was re-evaluated with an independent three-dimensional finite element solution. A seven-tooth model was built by rotating and arraying the discrete points of the tooth surface equations, so that both members satisfied the required surface accuracy, and the two members were assembled in a common meshing coordinate system. The mesh was exported in inp format and solved in ABAQUS with identical material properties for both members (elastic modulus 206 GPa, Poisson’s ratio 0.29, density 7800 kg/m³), frictionless hard surface-to-surface contact between the pinion concave working flank as master surface and the gear convex flank as slave surface, displacement couplings between each axis reference point and its tooth section, the pinion fully constrained and the torque applied to the gear with its rotational degree of freedom released. Post-processing converted the results into loaded transmission error along the mesh cycle.
The finite element and loaded tooth contact analysis curves differed slightly in amplitude but reproduced the same shape and the same ranking: the optimized parabolic design showed the larger ripple and the wave design the smaller one. This agreement supports the conclusion that the improvement is a property of the surface geometry rather than an artefact of the compliance model used inside the optimization loop.
Behaviour Across Load Conditions
Aircraft gears do not operate at a single torque, so the comparison was repeated over a range of loads. Across that range the unoptimized parabolic design gave the largest LTE amplitude and dynamic load factor, the optimized parabolic design was intermediate, and the wave design was lowest throughout. The persistence of the ranking matters more than any single value: it indicates that the concave target curve, whose coefficients were derived at the design load, does not lose its advantage when the deflection pattern changes, which is the practical requirement for a transmission that must run quietly over a mission profile.
Engineering Implications and Limits
The result quantifies a point that is often stated qualitatively: the shape of the transmission error function must match the contact-ratio regime of the gear pair. A correction that is optimal for ε < 2 cannot be scaled into effectiveness for ε > 2, because the number of stiffness transitions per mesh cycle differs. Raising the polynomial order of the target curve from two to seven, and letting the optimizer find the modified-roll coefficients that realize it, reduced the LTE amplitude of a high-contact-ratio aerospace pair to below one arcsecond at the rated load — a level at which the mesh excitation approaches the level of other transmission-path sources.
Several limits follow from the formulation. The polynomial coefficients define one target curve for one contact ratio and one design load; a different tooth count, face width or operating torque requires the curve to be re-derived and the optimization repeated, and the multi-load results, while favourable, do not replace that step. The optimization is only as good as the loaded tooth contact analysis embedded in it, since the compliance model sets the deflection the surface must compensate; the finite element comparison confirms the trend but not the absolute amplitude. Realizing the surface demands a computer numerical control hypoid generator capable of the higher-order modified-roll motions, and the correspondence between polynomial coefficients and machine axes must be exact, otherwise the surface departs from the ease-off target. Finally, the dynamic assessment used an eight-degree-of-freedom model with a time-varying mesh stiffness and neglected friction; in a real gearbox, shaft and casing flexibility and bearing clearance will modify the response, and avoidance of edge contact must be verified over the full load envelope rather than at the design point alone. Within these limits, wave tooth surface design gives a direct route from a vibration target to a manufacturable spiral bevel gear tooth surface, and the same reasoning is transferable to other gear types whose mesh stiffness varies periodically with the number of contacting pairs.