Spiral bevel gears carry high torque in a single mesh and run smoothly, which is why they dominate final drives in automotive axles and in aerospace transmissions. Their load capacity, however, is set less by the blank dimensions than by how the two flanks touch. A design that produces a small, well-controlled contact ellipse is robust against misalignment but leaves most of the flank idle; a design that brings the whole flank into contact can raise capacity substantially, provided the contact can be realised without undercut or interference.
Why point contact became the default
The classical design route for tapered-tooth spiral bevel gears is built on local conjugate contact. Its origin is mechanical rather than tribological: the generated motions available on a mechanical cradle-type machine are limited, so the pinion flank is deliberately produced as a surface that touches the gear flank at a point rather than along a line. The resulting pair is conjugate only in a first-order sense at the chosen reference point, with the mismatch controlled through the ease-off. Every subsequent design tool — local synthesis, transmission-error-controlled design, and more recently ease-off flank topography correction — inherits this premise.
The penalty is geometric. A theoretical point contact under load becomes an ellipse whose size depends on the elastic deflection and on the curvature mismatch across the flank. Contact stresses concentrate near the ellipse, and the regions of the flank outside it contribute nothing to load sharing. For a gear set sized by pitting or bending at the critical contact position, this is a direct loss of torque density.
Machine-setting compensation changes the premise. By re-identifying the machine settings that generate the pinion, the manufactured flank can be driven toward the flank that is truly conjugate to the gear — a line-conjugate surface. Once that is achievable, the design question shifts from “how small should the contact ellipse be” to “which basic blank parameters allow a full-face line contact that stays inside the working flank”. That is the problem addressed by the optimization below.
Constructing the line-conjugate pinion flank
Take the gear flank as known and generated. The pinion flank that is line-conjugate to it, denoted the theoretical flank $\Sigma_1$, follows from the meshing condition together with the coordinate transformation:
$$\mathbf{n}\cdot\mathbf{v}_{12}=0,\qquad \mathbf{r}_2=\mathbf{r}_2(u,\theta),\qquad \mathbf{r}_1=\mathbf{M}_{12}\,\mathbf{r}_2$$
where $\mathbf{n}$ is the unit normal of the gear flank, $\mathbf{v}_{12}$ the relative velocity of the pair, $(u,\theta)$ the gear flank parameters, and $\mathbf{M}_{12}$ the transform from the gear to the pinion coordinate system.
The flank actually cut on the pinion, $\Sigma_1^{*}$, is obtained from the machine settings $\mathbf{d}$ taken off the cutting-tool data sheet and inserted into the cutting model; it is point-conjugate to the gear flank by construction. To close the gap between $\Sigma_1^{*}$ and $\Sigma_1$, the pinion flank is meshed into $m$ grid points, and at each point $i$ the normal mismatch relative to the theoretical flank is evaluated:
$$h_i=\left[\mathbf{R}_i-\mathbf{R}_{i1}(\varepsilon_i,\mathbf{d})\right]\cdot\mathbf{n}_{i1}(\varepsilon_i,\mathbf{d})$$
Here $\mathbf{R}_i$ is the position vector of grid point $i$ on $\Sigma_1$, while $\mathbf{R}_{i1}$ and $\mathbf{n}_{i1}$ are the position and normal produced by the current settings at the grid coordinate $\varepsilon_i$. The set of $h_i$ over the grid forms the geometric topological error surface of the pinion flank, and the theoretical flank is recovered by superposing that error surface on the machined flank.
Correcting the settings is then a least-squares problem in $\mathbf{d}$:
$$\min_{\mathbf{d}} f(\mathbf{d})=\tfrac{1}{2}\,\mathbf{h}(\mathbf{d})^{\mathsf T}\mathbf{h}(\mathbf{d})$$
Because the mismatch is a nonlinear function of the settings, the solution is obtained iteratively with a Levenberg–Marquardt scheme under a trust-region strategy, updating $\mathbf{d}$ until the machined flank approaches the theoretical one within tolerance. Line contact is thus a manufacturing outcome as much as a design outcome: the flank geometry must permit it, and the settings must be identified well enough to realise it.
Keeping the contact zone inside the working flank
Full-face contact is admissible only if two characteristic curves stay off the working region. The first is the meshing limit line, along which the relative velocity becomes tangent to the flank and the contact condition degenerates:
$$\mathbf{r}_2=\mathbf{r}_2(u,\theta,\varphi),\qquad \mathbf{n}\cdot\mathbf{v}_{12}=0,\qquad \Phi_\varphi(u,\theta,\varphi)=0$$
$$\Phi_\varphi=\mathbf{n}\cdot\mathbf{q},\qquad \mathbf{q}=\boldsymbol{\omega}_1\times(\boldsymbol{\omega}_2\times\mathbf{r}_2)-\boldsymbol{\omega}_2\times(\boldsymbol{\omega}_1\times\mathbf{r}_1)$$
This line divides the gear flank into a region that can be reached by the contact and a region that cannot. The second is the root-cutting limit line, at which the generated pinion would be undercut or the two bodies would interfere:
$$\psi(u,\theta,\varphi)=\frac{1}{D^{\,2}}\begin{vmatrix}E_2 & F_2 & \mathbf{r}_{u2}\cdot\mathbf{v}_{12}\\ F_2 & G_2 & \mathbf{r}_{\theta2}\cdot\mathbf{v}_{12}\\ \Phi_u & \Phi_\theta & \Phi_\varphi\end{vmatrix}=0,\qquad D=E_2G_2-F_2^{\,2}$$
with $E_2$, $F_2$, $G_2$ the coefficients of the first fundamental form of the gear flank, $\mathbf{r}_{u2}$ and $\mathbf{r}_{\theta2}$ its partial derivatives, and $\Phi_u$, $\Phi_\theta$, $\Phi_\varphi$ the partial derivatives of the meshing function.
Both curves are tested in the axial section of the gear, where their rotational projections are compared with the root curve and the tip curve. Four equally spaced comparison points $P_j$ ($j=1,\dots,4$) are placed from the toe to the heel. For the meshing limit line, the projection radius $R_{fj}$ is compared with the root-curve radius $r_{fj}$; for the root-cutting limit line, $R_{aj}$ is compared with the tip-curve radius $r_{aj}$. Only positive violations are accumulated, so the constraint is exactly zero when the curve lies outside the flank on the correct side and positive otherwise:
$$h_1(\mathbf{x})=\sum_{j=1}^{4}\max\left(R_{fj}-r_{fj},\,0\right),\qquad h_2(\mathbf{x})=\sum_{j=1}^{4}\max\left(r_{aj}-R_{aj},\,0\right)$$
A third condition guards against the addendum becoming knife-edged. The pinion tip thickness $s_{a1}$ must satisfy
$$g(\mathbf{x})=s_{a1}-0.35\,\frac{m_n}{\cos\beta}\ \ge\ 0$$
with $m_n$ the normal module and $\beta$ the spiral angle. This bound is the one most sensitive to the profile shift coefficient, which is why the shift enters the design vector.
Two objectives: oil-film formation and contact-line length
Lubrication is assessed through an elastohydrodynamic film estimate. The minimum film thickness follows the Hamrock–Dowson relation
$$A_{\min}=\frac{2.65\,\alpha^{0.54}\left(\eta_0 u_0\right)^{0.7}}{R^{0.43}E^{0.03}Q^{0.13}}$$
where $\alpha$ is the pressure–viscosity coefficient, $\eta_0$ the dynamic viscosity at the operating temperature, $u_0$ the entrainment velocity, $R$ the equivalent radius of curvature at the mesh point, $E$ the equivalent elastic modulus, and $Q$ the contact load per unit width. Separating the geometry-dependent terms from the operating terms leaves a single comprehensive parameter,
$$v_\rho=u_0^{0.7}R^{0.43}$$
which is large when the surfaces entrain oil quickly and are generously curved — both favourable to film formation.
The equivalent radius follows from the contact-line normal $\mathbf{N}$ at the mesh point, computed from the fundamental-form coefficients and the meshing function derivatives, giving $R=\psi/|\mathbf{N}|^{2}$. The entrainment velocity is built from the two surface velocities and the direction of relative motion, $\mathbf{g}_{12}=\mathbf{v}_{12}/|\mathbf{v}_{12}|$:
$$u_0=\left|\frac{\mathbf{v}_1\cdot\mathbf{g}_{12}+\mathbf{v}_2\cdot\mathbf{g}_{12}}{2}\right|$$
Because the weakest point on the flank governs, the first objective is written against the minimum over the $n$ discretised grid points:
$$\min f_1(\mathbf{x})=\frac{1}{\min_i v_{\rho i}},\qquad i=1,\dots,n$$
The second objective addresses load sharing directly. The total length of the instantaneous contact lines at a given meshing instant measures how much flank is simultaneously engaged, so a longer total line means lower stress for the same torque. Taking the worst instant over the single-tooth mesh duration $T$:
$$\min f_2(\mathbf{x})=\frac{1}{\min_t l_t},\qquad t=0,\dots,T$$
with $l_t$ the summed length of all instantaneous contact lines at time $t$.
The penalised optimization model
The design vector collects the spiral angle $\beta$, the working tooth-height coefficient $h_k$, and the profile shift coefficient $\xi$:
$$\mathbf{x}=\left[\beta,\ h_k,\ \xi\right]^{\mathsf T}$$
$\beta$ governs the contact ratio and hence the engagement length; $h_k$ controls where the two limit lines fall relative to the working region; $\xi$ trades contact strength against tip sharpening. The constrained problem is converted to an unconstrained one by an exterior penalty, with the equality violations and the inequality violation weighted equally:
$$\min \mathbf{F}(\mathbf{x})=\left\{f_1(\mathbf{x}),\,f_2(\mathbf{x})\right\},\qquad f_{1,2}(\mathbf{x})=\frac{1}{\min v_\rho,\ \min l_t}+\sigma_{st}$$
$$\sigma_{st}=\sigma_1\sum_{b=1}^{2}h_b^{2}(\mathbf{x})+\sigma_2\max\left\{0,-g(\mathbf{x})\right\}^{2},\qquad \sigma_1=\sigma_2=1000$$
subject to $\beta\in[35°,40°]$, $h_k\in[1.5,1.8]$, $\xi\in[0.2,0.5]$.
The two objectives genuinely conflict. Large spiral angle and large working tooth height lengthen the contact lines and raise load capacity, but they also increase sliding and shift the limit lines toward the working region; the profile shift that strengthens the tooth root also thins the tip. There is no single optimum, only a Pareto set from which one design must be chosen.
Solving with a bare-bones multi-objective particle swarm
Gradient-based methods are unattractive here. The objective functions are composites of the meshing equations, and assembling the gradient matrix is laborious and error-prone. Population-based search avoids the derivatives entirely. The bare-bones multi-objective particle swarm variant used here dispenses with the velocity update and samples each particle’s new position from a Gaussian distribution centred between its personal best and the global best, which produces an evenly spread Pareto front with few tuning parameters.
The procedure meshes the gear and pinion flanks, evaluates the limit-line projections and the objective terms over the grid for each candidate design vector, applies the penalty, and updates the swarm until the termination criterion is met — 200 generations, with the front effectively stationary after about 100. The output is the Pareto set of basic parameters and the corresponding objective values.
Worked example: a 29/11 pair
The procedure was applied to a moderately high-speed ratio pair, 29 gear teeth against 11 pinion teeth, with the gear right-hand and the pinion left-hand, a 90° shaft angle and a 5 mm module. The gear was generated by the forming method and the pinion by the modified-roll method, with the pinion settings compensated so that its flank becomes line-conjugate to the gear flank. The gear cutter had a 76.2 mm radius with inner and outer blade profile angles of 22° and −18°.
| Parameter | Pinion | Gear |
|---|---|---|
| Number of teeth | 11 | 29 |
| Hand of spiral | left | right |
| Shaft angle | 90° | |
| Module | 5 mm | |
| Pitch cone angle | 20.77° | 69.23° |
| Outer cone distance | 77.54 mm | |
| Inner cone distance | 54.54 mm | |
| Face width | 23 mm | |
A balanced point on the Pareto front was selected rather than an extreme, giving the optimized design $\left[\beta^{*},h_k^{*},\xi^{*}\right]^{\mathsf T}=\left[35.07°,1.70,0.48\right]^{\mathsf T}$. The comparison baseline is the conventional value taken from design tables, $\left[35.00°,1.65,0.39\right]^{\mathsf T}$.
What changed on the flank
The geometric effect is qualitative before it is numerical. With the baseline parameters, a root-cutting limit line falls inside the gear flank: when the gear convex surface envelops the pinion concave surface, the heel region of the gear addendum penetrates the pinion body, producing undercut or solid interference. With the optimized parameters, no root-cutting limit line appears on the flank at all, and the two bodies mesh entirely outside each other. Over the whole admissible range of the design variables, the meshing limit line stayed outside the working region both before and after optimization, so it was not the binding constraint in this case; the root-cutting line was.
The quantitative gains are modest but consistent, and they are measured at the worst point of the flank, which is the dimensioning location:
| Design | $\beta$ / ° | $h_k$ | $\xi$ | $\min v_\rho$ | $\min l_t$ / mm |
|---|---|---|---|---|---|
| Baseline (table values) | 35.00 | 1.65 | 0.39 | 6.1542 | 20.3257 |
| Optimized | 35.07 | 1.70 | 0.48 | 6.4248 | 21.3472 |
| Change | — | — | — | +4.4% | +5.0% |
The lubrication parameter rises by 4.4% and the worst-instant contact-line length by 5.0%. Read against the baseline, roughly one twentieth more flank is engaged at the critical instant, and the minimum film estimate improves by a similar margin.
Finite element check
Both parameter sets were converted into three-dimensional models by solving the flank equations pointwise, then loaded in a finite element contact analysis. The baseline model shows interference at the boundary between the gear addendum at the heel and the pinion root, with the attendant stress concentration — the numerical signature of the root-cutting limit line lying inside the flank. The optimized model shows the working flanks in line contact across the full face, with no interference and no limit-line influence inside the working region. The finite element result therefore reproduces the geometric prediction rather than adding independent evidence: both follow from the same flank equations.
Scope and limitations
Three points bound the claims that can be made for this approach. First, the improvement is a few percent, not a step change, and it is achieved by moving three blank parameters inside narrow bounds; the contact-line length cannot grow without limit because the limit lines eventually migrate into the working region, and it is that migration, not the objective, that sets the ceiling. Second, the assessment is purely geometric and elastohydrodynamic: film thickness is estimated from the Hamrock–Dowson formula with nominal operating conditions, contact stress is examined only qualitatively through the finite element model, and no loaded tooth contact analysis or rig test is reported, so the predicted gains in film and load sharing remain unconfirmed experimentally. Third, the conclusion that the meshing limit line is non-binding holds for this tooth ratio, this cutter geometry and this variable window; a different ratio or a wider window could make it the active constraint.
What the method does provide is a systematic route from a manufacturability constraint to a design decision. Local conjugate point contact was a consequence of machine-tool kinematics, not of gear theory; once setting compensation makes line conjugation reachable, the blank parameters can be optimized against the two quantities that actually govern flank performance — the worst-point film-forming parameter and the worst-instant total contact-line length — while the limit lines and the tip-thickness bound are held as explicit constraints. The result is a design procedure that runs entirely on flank geometry, requires no gradient information, and yields a Pareto set from which a manufacturing-constrained choice can be made.