Spiral bevel gears transmit torque between intersecting shafts through a curved tooth trace. Their flanks are deliberately crowned so that unloaded contact is confined to a small ellipse that travels across the tooth during mesh, rather than spreading over the full face. The size, orientation, and path of that ellipse, together with the variation in output rotation, determine load sharing, noise, sensitivity to mounting error, and the risk of edge contact. Tooth contact analysis (TCA) is the standard way to predict these quantities, but a TCA result is only as good as the machine-tool settings used to generate the surfaces. A conventional design loop varies cutter radius, pressure angle, radial and angular cutter positions, roll ratio, bed position, wheel positions, and cutter tilt. The map from those settings to contact quality is strongly coupled and difficult to interpret. TCA is also normally performed at light load and does not by itself impose a target on transmission error.
A more compact formulation is possible. At the pitch point, the pinion generating surface can be described by three local quantities: normal curvature in the tooth-height direction, normal curvature in the tooth-length direction, and geodesic torsion. These three variables determine the local meshing geometry and therefore the contact ellipse and transmission-error curve. The design problem becomes a search in three curvature variables instead of a manual adjustment of many machine settings. A genetic algorithm is suitable for this search because the objective is non-smooth and the feasible contact pattern is bounded by geometric constraints.
Contact Geometry as a Reduced-Order Design Problem
At the pitch point \(P\), where the pitch cones of the gear and pinion are tangent, the local shape of the pinion tooth surface is fixed by the normal curvature in the tooth-height direction \(A_P\), the normal curvature in the tooth-length direction \(B_P\), and the geodesic torsion \(C_P\). These quantities determine the orientation and spacing of the instantaneous contact ellipses in the neighborhood of \(P\). Blank parameters such as spiral angle, pitch cone angle, and pitch cone distance set the overall size and cone geometry, but they do not by themselves define the tooth form. They are therefore poor search variables when the goal is contact correction. The curvature triplet is the natural control vector because meshing near \(P\) is governed primarily by the relative principal curvatures and torsion of the two surfaces.
The choice also reduces dimensionality. Instead of asking which of many machine settings should change, the designer asks what local surface shape should be produced; the machine settings are then calculated as the realization of that shape. The optimization is a pre-design step, not a replacement for the machine-setting calculation.
Three Contact Targets
Contact ellipse size. The semi-major axis \(l\) of the instantaneous contact ellipse controls the contact area. It should be large enough to keep Hertzian pressure and misalignment sensitivity low, but not so large that the ellipse approaches the tooth edges and produces edge loading. A target value is
$$l_{\mathrm{OPT}}=\frac{F b}{2\cos\beta},$$
where \(F\) is a face-width contact coefficient, \(b\) is face width, and \(\beta\) is spiral angle. Because many ellipses occur in one mesh cycle and their lengths differ, the optimization uses the mean
$$\bar l=\frac{1}{n}\sum_{k=0}^{n-1}l_k,$$
with \(n\) the number of instantaneous ellipses in the cycle. An ellipse that is too short concentrates stress; one that is too long risks edge loading and a contact pattern that is difficult to control under misalignment.
Contact trace direction. The angle \(\gamma\) is measured between the contact path and the root-cone generatrix. It couples the contact pattern to the shape of the transmission-error curve. For a convex gear flank, \(\gamma<0\) tends to separate the transmission-error branches, while \(\gamma>0\) makes adjacent branches overlap. Increasing \(\gamma\) raises the contact ratio but eventually produces severe internal diagonal contact and demands high mounting accuracy. For a concave flank the sign is reversed. The design windows used are
$$\gamma_{\mathrm{OPT}}\in[50^\circ,70^\circ]\ \text{(convex gear flank)},\qquad
\gamma_{\mathrm{OPT}}\in[110^\circ,130^\circ]\ \text{(concave gear flank)}.$$
The trace is fitted linearly,
$$P_{ay}(P_{ax})=\sum_{i=0}^{1}c_i(P_{ax})^i,\qquad \gamma=\arctan c_1,$$
so the angle is extracted from the computed contact points rather than assumed from the blank geometry.
Transmission-error intersection ordinate. The ordinate \(\delta\) of the intersection of adjacent transmission-error branches controls whether the motion error is continuous and whether the curve bends in the intended direction. It should be negative and close to zero so that the actual ratio stays near the theoretical ratio, but not so close that the curve bends upward or loses a stable intersection. The target window is
$$\delta_{\mathrm{OPT}}\in[-6\times10^{-5},-4\times10^{-5}]\ \mathrm{rad}.$$
A fourth-order polynomial is fitted to the transmission error,
$$\Delta\psi(\psi_g)=\sum_{i=0}^{4}c_i\psi_g^i,$$
and the ordinate at the crossing is
$$\delta=c_0+c_2\left(\frac{\pi}{N_g}\right)^2+c_4\left(\frac{\pi}{N_g}\right)^4,$$
with \(N_g\) the gear tooth count. The coefficients are not free design variables; they are consequences of the curvature triplet and are recomputed for each candidate.
Combined Objective, Constraints, and Genetic Search
The three targets are normalized and combined:
$$
f(l,\gamma,\delta)=
\max\left\{\left|\frac{l-l_{\mathrm{OPT}}}{l_{\mathrm{OPT}}}\right|,\Delta\varepsilon\right\}
+\max\left\{\left|\frac{\gamma-\gamma_{\mathrm{OPT}}}{\gamma_{\mathrm{OPT}}}\right|,\Delta\varepsilon\right\}
+\max\left\{\left|\frac{\delta-\delta_{\mathrm{OPT}}}{\delta_{\mathrm{OPT}}}\right|,\Delta\varepsilon\right\},
$$
where \(\Delta\varepsilon\) is the acceptable relative tolerance. When each normalized deviation is at or below \(\Delta\varepsilon\), \(f=3\Delta\varepsilon\); a larger value indicates that at least one target is outside tolerance. Normalization is necessary because \(l\) is in millimeters, \(\gamma\) in degrees, and \(\delta\) in radians. Without it, the search would be dominated by the numerically largest term.
A candidate is rejected if the contact point falls outside either tooth surface. For the gear and pinion, the four triangular areas bounded by the contact point and the top, toe, root, and heel lines must sum to the corresponding axial-section area:
$$
\sum_{q\in\{\mathrm{top},\mathrm{toe},\mathrm{root},\mathrm{heel}\}}\Delta S_q^{g}=\Delta S_g,\qquad
\sum_{q\in\{\mathrm{top},\mathrm{toe},\mathrm{root},\mathrm{heel}\}}\Delta S_q^{P}=\Delta S_P.
$$
If a sum exceeds the surface area, the point lies outside the tooth and the candidate is removed before fitness evaluation.
Control variables are binary-coded. Fitness is \(fit=-\log f\), so smaller objective values receive higher fitness. Selection uses roulette probabilities
$$P_i=\frac{fit_i}{\sum_{j=1}^{m}fit_j},$$
with population \(m=30\), crossover probability 0.6, mutation probability 0.09, and a maximum of 250 generations. These are modest settings; the purpose is reliable convergence in a three-variable space rather than exhaustive global search.
The converged convex-flank solution used the targets \(l_{\mathrm{OPT}}=3.0036\) mm, \(\gamma_{\mathrm{OPT}}=59.9693^\circ\), and \(\delta_{\mathrm{OPT}}=-5.00077\times10^{-5}\) rad, with curvature parameters \(A_P=-0.0026\), \(B_P=0.031988\), \(C_P=-0.01723\), and tolerance \(\Delta\varepsilon=0.001\). The objective reached \(3\Delta\varepsilon\), so the computed contact metrics were within 0.1% of the target values. Table 1 collects the target and curvature values.
| Quantity | Symbol | Value |
|---|---|---|
| Contact ellipse semi-major axis | \(l_{\mathrm{OPT}}\) | 3.0036 mm |
| Contact trace direction angle | \(\gamma_{\mathrm{OPT}}\) | 59.9693° |
| Transmission-error intersection ordinate | \(\delta_{\mathrm{OPT}}\) | \(-5.00077\times10^{-5}\) rad |
| Relative tolerance | \(\Delta\varepsilon\) | 0.001 |
| Normal curvature, tooth-height direction | \(A_P\) | −0.0026 |
| Normal curvature, tooth-length direction | \(B_P\) | 0.031988 |
| Geodesic torsion | \(C_P\) | −0.01723 |
| Objective at convergence | \(f\) | \(3\Delta\varepsilon\) |
The curvature solution is realized by modest changes to the pinion machine settings. Table 2 lists the corrections for the convex and concave flanks. The corrections are not symmetric: several settings reverse sign, reflecting the different sign of the required surface correction on the two sides of the tooth. The largest translational corrections are about 2.6 mm in axial and vertical wheel position and about 2.1 mm in radial cutter position. Angular corrections remain below about 2.2° in pressure angle and below 0.6° in most other angular settings.
| Parameter | Convex flank | Concave flank |
|---|---|---|
| Cutter radius \(\Delta r_{0P}\) (mm) | −0.8849 | 0.8385 |
| Cutter pressure angle \(\Delta\alpha’_P\) (°) | 2.1312 | −1.9527 |
| Cutter top distance \(\Delta U_P\) (mm) | — | — |
| Radial cutter position \(\Delta V_P\) (mm) | 2.0619 | −2.1729 |
| Angular cutter position \(\Delta q_P\) (°) | −0.5621 | 0.4975 |
| Machine roll ratio \(\Delta i_P\) | 0.1129 | −0.0109 |
| Bed position \(\Delta Y_{BP}\) (mm) | 0.2037 | −0.0826 |
| Axial wheel position \(\Delta Y_P\) (mm) | 2.65 | −2.6435 |
| Vertical wheel position \(\Delta F_P\) (mm) | 2.54 | −2.54 |
| Blank mounting angle \(\Delta\Gamma_{MP}\) (°) | 0.1731 | — |
| Cutter tilt angle \(\Delta i\) (°) | 0.5234 | −0.1831 |
| Cutter swivel angle \(\Delta j\) (°) | 0.2893 | −0.57 |
These values are corrections to the nominal settings, not absolute machine coordinates. They define the cutting geometry that reproduces the optimized curvature triplet.
Finite-Element Verification Under Load
The optimized geometry was checked with a three-dimensional finite-element model of the gear pair in ANSYS. The contact ratio is 1.6134, so the mesh was simplified to two tooth pairs and the contacting flanks were locally refined within a mapped surface mesh. The unloaded optimization cannot by itself show how the contact ellipse moves under torque, so the model was run at 100, 200, and 400 Nm.
On the convex gear flank, the contact pattern remained basically elliptical and centered on the flank, consistent with the zero-load target. At 100 Nm the von Mises stress reached 241.9 MPa at the center and fell to 2.016 MPa near the root. At 200 Nm the peak was 318.3 MPa and the root-region value rose to 7.959 MPa. At 400 Nm the root-region value reached 14.93 MPa; the contact area and trace length increased, and the trace direction angle decreased. The direction of migration is the important point: the pattern grew toward the root, not toward the toe. This means the full face width can be used under load if the light-load pattern is placed near the toe during assembly. A mounting position that centers the light-load ellipse may push the loaded ellipse too far toward the root.
The stress distribution was smooth, and the loaded pattern did not develop the diagonal or edge-contact forms that the optimization was intended to suppress. These results are consistent with the optimized zero-load pattern, but they do not establish fatigue life or scoring resistance; those require surface integrity, lubrication, and test data.
Transmission Error and Instantaneous Ratio
Transmission error is the deviation of the output rotation from the ideal ratio. Table 3 summarizes the computed values. The theoretical curve has a peak-to-peak error of \(5.743\times10^{-5}\) rad. Under 100 Nm the error ranges from \(3.055\times10^{-8}\) to \(-2.351\times10^{-5}\) rad, with a peak-to-peak value of \(2.3541\times10^{-5}\) rad; the curve has three fluctuation zones per cycle, corresponding to switching between single and double tooth contact. At 200 Nm the range is \(-2.973\times10^{-8}\) to \(-3.279\times10^{-5}\) rad and the peak-to-peak value is \(3.260\times10^{-5}\) rad, again with three zones. At 400 Nm the range widens to \(-2.053\times10^{-7}\) to \(-5.431\times10^{-5}\) rad, with a peak-to-peak value of \(5.4105\times10^{-5}\) rad. The distinct three-zone pattern disappears because load sharing spreads across more teeth, although the peak-to-peak error is the largest of the three load cases.
| Load | Max \(\Delta\varepsilon\) (rad) | Min \(\Delta\varepsilon\) (rad) | Peak-to-peak (rad) | Instantaneous-ratio deviation | Mesh behavior |
|---|---|---|---|---|---|
| Theoretical | \(-4.64\times10^{-11}\) | \(-5.743\times10^{-5}\) | \(5.743\times10^{-5}\) | — | Periodic |
| 100 Nm | \(3.055\times10^{-8}\) | \(-2.351\times10^{-5}\) | \(2.3541\times10^{-5}\) | 0.006 | Three zones; single/double contact |
| 200 Nm | \(-2.973\times10^{-8}\) | \(-3.279\times10^{-5}\) | \(3.260\times10^{-5}\) | 0.006 | Three zones; single/double contact |
| 400 Nm | \(-2.053\times10^{-7}\) | \(-5.431\times10^{-5}\) | \(5.4105\times10^{-5}\) | 0.009 | Smoother; double/triple contact |
The instantaneous ratio stayed near the theoretical 2.083. Deviations were 0.006 at 100 and 200 Nm and 0.009 at 400 Nm. The increase with torque indicates that kinematic steadiness deteriorates even as the contact ratio rises. Load sharing smooths the local waveform but does not remove the periodic error.
Design Implications and Limits
The value of the method lies in reducing a coupled machine-setting problem to a three-variable curvature problem with explicit targets. It makes the trade-off among contact size, trace angle, and transmission-error intersection visible, and it provides a convergence criterion rather than a visual judgment of the contact pattern. The genetic algorithm handles the non-smooth feasible boundary and the coupling among the three targets. The finite-element pattern and transmission-error trends are consistent with the optimized targets, which is the main evidence for the approach.
The scope is nevertheless specific. The result applies to one gear pair, ideal geometry, torques up to 400 Nm, and a light-load TCA basis. It does not include manufacturing scatter, thermal distortion, lubrication, friction, or misalignment beyond the modeled mounting condition. The machine-setting corrections are tied to the nominal cutter and machine. The optimization is therefore a pre-design and a hypothesis for test, not a guarantee of service performance. Within these limits, the combination of curvature parameterization, three contact targets, and genetic search gives a coherent route from tooth-surface shape to loaded behavior.