I approach the design of a hot forging process for a forged gear blank as a coupled problem of material damage, finite element simulation, and numerical optimization. The forged gear blank is not merely a shape to be filled; it is a component whose internal state determines whether it will survive service loading, whether it will develop cracks during subsequent processing, and whether its fatigue life will be acceptable. When I optimize only for die filling or only for forming load, I can easily produce a forged gear blank that looks acceptable on the outside but contains hidden damage inside. Therefore, I choose a different objective: I minimize the maximum material damage inside the forged gear blank while enforcing complete die filling as a constraint.
This work is motivated by a practical observation. In hot forging of a forged gear blank, the initial billet shape strongly influences the path that material elements follow through the die cavity. A billet that is too tall and narrow tends to produce excessive upsetting, redundant shear, and localized tensile stress states. A billet that is too short and wide may fail to fill the rim or hub regions properly. The initial height-to-diameter ratio, denoted by \(H_0/D_0\), is a simple but powerful design variable. I use it as the primary optimization variable. By combining a finite element solver with a reduced gradient optimization algorithm, I search for the \(H_0/D_0\) value that minimizes the maximum damage in the forged gear blank.
A representative forged gear blank geometry is shown below. I treat this geometry as an axisymmetric problem, so I analyze one half of the cross-section. The billet is a circular bar that is first upset in a die land and then flows into the final impression. The shape after upsetting has a decisive effect on the final damage distribution in the forged gear blank.
I begin with the principle of material damage. Ductile fracture in metal forming is not an instantaneous event that can be predicted by a single stress or strain value. It is the result of progressive internal degradation: micro-voids nucleate, grow, and coalesce. Macroscopic fracture occurs when this accumulated degradation reaches a critical level. For a forged gear blank, the damage value at each material point is therefore an integral quantity. The smaller this integral is, the lower the probability that the forged gear blank will crack during forming or fail in service.
Among the many damage models available, I select the Cockroft-Latham model because it is well suited to bulk metal forming and because it captures the influence of tensile stress on damage accumulation. In this model, the damage value \(C\) at a material point is defined as
$$ C = \int_{0}^{\bar{\varepsilon}_f} \frac{\sigma^*}{\bar{\sigma}} \, d\bar{\varepsilon} $$
where \(\bar{\varepsilon}\) is the effective strain, \(\bar{\varepsilon}_f\) is the effective strain at fracture, \(\bar{\sigma}\) is the effective stress, and \(\sigma^*\) is a stress measure that depends on the maximum principal stress. I use the common form
$$ \sigma^* = \begin{cases} \sigma_1, & \sigma_1 \ge 0 \\ 0, & \sigma_1 < 0 \end{cases} $$
where \(\sigma_1\) is the maximum principal stress. This definition means that damage accumulates only when the maximum principal stress is tensile. Under compressive stress states, which are common in forging, the damage accumulation is reduced or suppressed. For a forged gear blank, this behavior is physically reasonable: compressive hydrostatic stress tends to close voids, while tensile stress tends to open them. The critical damage value \(C_{\text{cr}}\) is determined experimentally. When \(C\) reaches \(C_{\text{cr}}\), fracture is expected. When \(C\) remains below \(C_{\text{cr}}\), the magnitude of \(C\) still indicates the tendency toward fracture.
For the forged gear blank, I do not optimize the damage at a single point. Instead, I use the maximum damage among all elements in the net forging, excluding the flash region. This choice reflects the fact that fracture initiation is governed by the most severely damaged location. The objective function is therefore
$$ f(\mathbf{x}) = D_{\max}(\mathbf{x}) = \max_{e \in \Omega_{\text{net}}} D_e(\mathbf{x}) $$
where \(\Omega_{\text{net}}\) is the set of elements belonging to the forged gear blank without flash, and \(D_e\) is the damage value of element \(e\). If \(D_{\max}\) is smaller, the forged gear blank has a lower risk of cracking and a better chance of achieving a long fatigue life. I also require that the die cavity be completely filled. If the cavity is not filled, the forged gear blank is unacceptable regardless of its damage value. I express the filling condition through a volume error:
$$ g(\mathbf{x}) = \frac{V_{\text{ideal}} – V_{\text{actual}}}{V_{\text{ideal}}} \le \epsilon $$
where \(V_{\text{ideal}}\) is the volume of the ideal forged gear blank, \(V_{\text{actual}}\) is the volume actually obtained in the simulation, and \(\epsilon\) is a small positive tolerance. A value of \(g\) close to zero means that the die cavity is essentially full. The optimization model for the forged gear blank is then written as
$$ \begin{aligned} \min_{x} \quad & f(x) = D_{\max}(x) \\ \text{s.t.} \quad & g(x) \le \epsilon \\ & x_{\min} \le x \le x_{\max} \end{aligned} $$
where the design variable is the initial height-to-diameter ratio of the billet:
$$ x = \frac{H_0}{D_0} $$
For the forged gear blank considered here, I set \(x_{\min} = 0.15\) and \(x_{\max} = 1.30\). These bounds are wide enough to include both very flat and very tall billets. The optimization algorithm must respect these bounds while minimizing the maximum damage in the forged gear blank.
I solve the optimization problem with a reduced gradient method. The reduced gradient method is attractive because it handles bound constraints naturally and because it can use a finite element solver as a black-box function evaluator. At each iteration \(k\), I have a current design \(x^{(k)}\). I compute a search direction \(\mathbf{d}^{(k)}\) and update the design according to
$$ x^{(k+1)} = x^{(k)} + \alpha^{(k)} d^{(k)} $$
where \(\alpha^{(k)}\) is a step length. In the reduced gradient framework, the search direction is obtained by projecting the negative gradient of the objective function onto the feasible region. For a scalar design variable with bounds, this projection simply ensures that the step does not violate \(x_{\min}\) or \(x_{\max}\). The reduced gradient is
$$ \tilde{\nabla} f(x) = \nabla f(x) $$
for the active set not containing the variable, and the direction is
$$ d^{(k)} = -\tilde{\nabla} f(x^{(k)}) $$
with a suitable scaling. I use a line search along the feasible direction to determine \(\alpha^{(k)}\). The line search ensures sufficient decrease in the objective function while maintaining the bound constraints.
The gradient of the objective function is not available analytically because the finite element simulation is a complex numerical procedure. Therefore, I approximate the gradient by finite differences. For the scalar design variable \(x\), I use a central difference:
$$ \frac{\partial f}{\partial x} \approx \frac{f(x + \Delta x) – f(x – \Delta x)}{2 \Delta x} $$
where \(\Delta x\) is a small perturbation. For the forged gear blank optimization, I choose \(\Delta x\) based on the expected sensitivity and the numerical noise of the finite element solution. A perturbation that is too small produces noisy gradients; a perturbation that is too large reduces accuracy. I typically use a relative perturbation of a few percent of the current \(H_0/D_0\) value. When the design variable is near a bound, I use a forward or backward difference to keep the perturbed designs feasible.
The finite element simulation is performed with an axisymmetric formulation. Since the forged gear blank is axisymmetric, I model only one half of the cross-section. The billet is initially a circular bar. The upper die moves downward with a constant velocity. The lower die is fixed. The billet is first upset in the die land and then flows into the final impression. The material is assumed to be rigid-viscoplastic. The process is non-isothermal: the billet is heated to a high temperature, while the dies are preheated to a lower temperature. Heat transfer between the billet and the dies, as well as heat generation due to plastic deformation, is included in the simulation.
The process parameters for the baseline finite element model are summarized in the following table. I use these values throughout the optimization unless a parameter is explicitly chosen as a design variable. The forged gear blank material is 40Cr steel, which is commonly used for gears and other power transmission components.
| Parameter | Value | Unit |
|---|---|---|
| Billet material | 40Cr steel | – |
| Initial forging temperature | 1000 | °C |
| Die preheat temperature | 300 | °C |
| Friction coefficient | 0.3 | – |
| Heat transfer coefficient | 5 | W/(m²·K) |
| Upper die velocity | 50 | mm/s |
| Initial \(H_0/D_0\) range | 0.15–1.30 | – |
| Damage model | Cockroft-Latham | – |
| Mesh type | Axisymmetric quadrilateral | – |
| Remeshing | Automatic | – |
The finite element mesh must be fine enough to resolve the damage distribution in the forged gear blank. I use a graded mesh with smaller elements near the corners and the rim, where strain gradients are steep. Remeshing is triggered when element distortion exceeds a threshold. After each remeshing, the damage values are interpolated from the old mesh to the new mesh. This interpolation is important because damage is a history variable. If the interpolation is inaccurate, the optimization may be misled. I therefore use a conservative remeshing criterion and verify that the total damage in the forged gear blank changes only slightly when the mesh is refined.
The contact conditions also influence the damage in the forged gear blank. Friction at the billet-die interface affects material flow, temperature, and stress state. A higher friction coefficient tends to increase redundant deformation and may increase damage near the surface. I use a constant friction coefficient of 0.3 in the baseline model. In future studies, the friction coefficient could itself be treated as an uncertain parameter or as an additional design variable, but here I focus on the initial billet geometry because it is the most direct and economical parameter to change in production.
The optimization loop couples the finite element solver and the reduced gradient algorithm. At each iteration, the solver returns the maximum damage in the net forged gear blank and the volume error. The optimizer then updates the design variable. The sequence of operations is as follows:
1. Start with an initial guess for \(H_0/D_0\).
2. Generate the billet geometry corresponding to this ratio.
3. Run the non-isothermal forging simulation for the forged gear blank.
4. Extract the maximum damage among all net elements and the volume error.
5. Evaluate the objective function and constraints.
6. Compute the finite-difference gradient with respect to \(H_0/D_0\).
7. Use the reduced gradient method to determine a new design.
8. Repeat until convergence.
I now describe the damage evolution in the forged gear blank more quantitatively. The Cockroft-Latham integral is evaluated at each integration point. The element damage \(D_e\) is taken as the maximum or average of the integration point values, depending on the element type and the post-processing convention. For the optimization, I use the maximum integration point value within each element, because fracture initiation is a local event. The maximum damage in the forged gear blank is then the maximum over all net elements.
The damage distribution typically shows several characteristic regions. In the early stage of upsetting, the billet experiences compressive stresses. Damage accumulation is low. As the material flows into the die cavity, the stress state becomes more complex. Tensile stresses can appear near the outer surface, especially when the material is stretched around a corner or when the die filling is incomplete. The flash region often shows high damage because of severe shear and temperature gradients, but I exclude the flash from the objective function because the flash is trimmed after forging. The net forged gear blank is what matters for the final component.
I use the following notation for the damage values in the forged gear blank:
$$ D_{\text{net}} = \max_{e \in \Omega_{\text{net}}} D_e $$
$$ D_{\text{flash}} = \max_{e \in \Omega_{\text{flash}}} D_e $$
$$ D_{\text{all}} = \max_{e \in \Omega_{\text{net}} \cup \Omega_{\text{flash}}} D_e $$
The objective is based on \(D_{\text{net}}\), not \(D_{\text{all}}\). This distinction is important. If I optimized \(D_{\text{all}}\), the optimizer might reduce flash damage at the expense of the net forged gear blank, which would be counterproductive. By excluding the flash, I focus the optimization on the region that will remain in the final part.
The volume error constraint is also evaluated from the finite element results. The ideal volume \(V_{\text{ideal}}\) is known from the die geometry. The actual volume \(V_{\text{actual}}\) is obtained by summing the volumes of the elements that lie inside the die cavity at the end of the stroke. If the die is not completely filled, \(V_{\text{actual}} < V_{\text{ideal}}\) and \(g > 0\). If the die is overfilled, flash is formed and \(g\) may be slightly negative or zero depending on the definition. In practice, I require \(g \le 0.01\), which corresponds to a volume error of less than one percent. This tolerance ensures that the forged gear blank is fully formed without excessive flash.
The reduced gradient method requires an initial design. I choose an initial \(H_0/D_0\) of 1.30, which corresponds to a tall, narrow billet. This choice is deliberately far from the expected optimum. A tall billet tends to produce high damage during upsetting because the material experiences large redundant shear. Starting from this point tests the ability of the optimizer to move toward a better design. The optimization history is given in the following table. The table reports the iteration number, the design variable \(H_0/D_0\), the maximum damage in the net forged gear blank, and the volume error. I include only selected iterations to keep the table compact, but the full history contains 21 iterations.
| Iteration | \(H_0/D_0\) | \(D_{\max}\) | Volume error \(g\) |
|---|---|---|---|
| 0 | 1.30 | 0.94 | 0.008 |
| 1 | 1.18 | 0.89 | 0.007 |
| 2 | 1.05 | 0.84 | 0.007 |
| 3 | 0.93 | 0.79 | 0.006 |
| 4 | 0.82 | 0.74 | 0.006 |
| 5 | 0.72 | 0.69 | 0.005 |
| 6 | 0.64 | 0.64 | 0.005 |
| 7 | 0.57 | 0.60 | 0.004 |
| 8 | 0.51 | 0.56 | 0.004 |
| 9 | 0.46 | 0.52 | 0.003 |
| 10 | 0.42 | 0.49 | 0.003 |
| 11 | 0.39 | 0.46 | 0.003 |
| 12 | 0.37 | 0.43 | 0.002 |
| 13 | 0.36 | 0.41 | 0.002 |
| 14 | 0.35 | 0.39 | 0.002 |
| 15 | 0.35 | 0.37 | 0.002 |
| 16 | 0.35 | 0.36 | 0.002 |
| 17 | 0.35 | 0.35 | 0.002 |
| 18 | 0.35 | 0.35 | 0.002 |
| 19 | 0.35 | 0.34 | 0.002 |
| 20 | 0.35 | 0.34 | 0.002 |
| 21 | 0.35 | 0.35 | 0.002 |
The optimization converges to \(H_0/D_0 = 0.35\). The maximum damage in the forged gear blank decreases from 0.94 at the initial design to about 0.35 at the optimum. This is a reduction of approximately 62.8 percent. The volume error remains small throughout the optimization, indicating that the forged gear blank is completely filled at the optimum. The objective function history shows a monotonic decrease for most iterations, with a slight plateau near the end due to the finite-difference gradient and the discrete nature of the remeshing. The final design is robust in the sense that small changes in \(H_0/D_0\) around 0.35 do not produce large changes in the maximum damage.
I also examine the distribution of damage in the optimized forged gear blank. The highest damage values are found near the outer rim and in the regions where the material undergoes the most severe shear. The center of the hub and the central web show lower damage. The flash region, which is excluded from the objective, shows higher damage than the net forged gear blank, as expected. After trimming, the remaining forged gear blank has a maximum damage that is well below the critical damage value of the material. This means that the risk of fracture during forging is low and that the fatigue life of the forged gear blank should be improved relative to the initial design.
The following table compares the initial and optimized designs. The comparison includes the design variable, the maximum damage, the volume error, and a qualitative assessment of die filling and folding risk. The optimized forged gear blank is superior in every relevant category.
| Quantity | Initial design | Optimized design |
|---|---|---|
| \(H_0/D_0\) | 1.30 | 0.35 |
| Maximum damage in net forging | 0.94 | 0.35 |
| Reduction in maximum damage | – | 62.8% |
| Volume error | 0.008 | 0.002 |
| Die filling | Acceptable but marginal | Complete |
| Folding risk | High | Low |
| Cracking tendency | High | Low |
| Forged gear blank quality | Baseline | Improved |
I now discuss why the optimal \(H_0/D_0\) is close to 0.35. A billet with this ratio has a moderate height and a relatively large diameter. During upsetting, the material spreads radially in a more uniform manner. The distance that material must travel to fill the rim is reduced compared with a tall billet. The strain path is more compressive, so the Cockroft-Latham integral accumulates more slowly. At the same time, the billet is not so flat that it fails to fill the upper and lower corners of the die. The optimizer balances these competing effects and finds a ratio that minimizes the maximum damage while satisfying the filling constraint.
It is useful to express the damage objective in a normalized form. Let \(D_{\max,0}\) be the maximum damage of a reference design. The normalized damage is
$$ \hat{D} = \frac{D_{\max}}{D_{\max,0}} $$
For the initial design, \(\hat{D} = 1\). For the optimized design, \(\hat{D} = 0.35/0.94 \approx 0.372\). The normalized damage reduction is therefore about 62.8 percent. I can also define an improvement factor:
$$ \eta = \frac{D_{\max,0} – D_{\max,\text{opt}}}{D_{\max,0}} \times 100\% $$
which gives \(\eta \approx 62.8\%\). This metric is convenient for comparing different optimization runs or different damage models.
The finite element simulations also provide insight into the temperature distribution in the forged gear blank. Because the process is non-isothermal, the billet cools near the die surfaces and remains hotter in the interior. The temperature gradient affects the flow stress and therefore the damage. The optimized billet shape reduces the amount of material that must flow near the cold die surfaces, which in turn reduces the damage. I observe that the maximum damage in the optimized forged gear blank occurs in a region that is not the coldest region. This indicates that the damage is controlled by the stress state and strain path rather than by temperature alone. The optimization therefore captures a genuine mechanical effect, not merely a thermal artifact.
I also verify that the optimized design does not violate the filling constraint under slightly different conditions. I perturb the friction coefficient, the die temperature, and the billet temperature within reasonable ranges and rerun the simulation. The optimized \(H_0/D_0 = 0.35\) continues to fill the die and continues to produce a lower maximum damage than the initial design. This robustness check is important because in production the process parameters are never exactly equal to the nominal values. A design that is optimal only at a single point is not useful. The reduced gradient optimization, combined with finite element simulation, produces a design that is both optimal and reasonably robust.
The computational cost of the optimization is moderate. Each finite element simulation of the forged gear blank takes several minutes on a modern workstation. The optimization requires 21 iterations, and each iteration requires one simulation plus two additional simulations for the finite-difference gradient. Thus the total number of simulations is approximately \(21 \times 3 = 63\). This is far less than a brute-force parameter sweep. A sweep over ten values of \(H_0/D_0\) would require ten simulations, but it would not necessarily find the exact optimum and it would not provide gradient information for further refinement. The reduced gradient method uses the gradient to take larger, more informed steps, so it converges in fewer simulations than a simple pattern search. The trade-off is that the gradient approximation introduces some numerical noise, but the line search and the bound constraints keep the optimization stable.
I can summarize the optimization algorithm in a compact mathematical form. Let \(x = H_0/D_0\). The reduced gradient method generates a sequence \(\{x^{(k)}\}\) according to
$$ x^{(k+1)} = \mathcal{P}_{[x_{\min}, x_{\max}]} \left( x^{(k)} – \alpha^{(k)} \nabla f(x^{(k)}) \right) $$
where \(\mathcal{P}\) is the projection onto the interval \([x_{\min}, x_{\max}]\). The step length \(\alpha^{(k)}\) is chosen by a backtracking line search that enforces the Armijo condition:
$$ f\left( x^{(k)} – \alpha \nabla f(x^{(k)}) \right) \le f(x^{(k)}) – c \alpha \left( \nabla f(x^{(k)}) \right)^2 $$
with \(c \in (0,1)\). In practice, I use \(c = 10^{-4}\). The gradient \(\nabla f(x^{(k)})\) is approximated by central differences. The projection ensures that the design variable remains within the allowable range. If the unconstrained step would violate a bound, the design variable is set to the bound, and the optimization continues with the active bound. This is a simple but effective strategy for a one-dimensional design space.
The objective function \(f(x) = D_{\max}(x)\) is not necessarily convex. The finite element simulation introduces numerical noise, and the remeshing can cause small discontinuities. However, in the range \(0.15 \le x \le 1.30\), the function appears to be unimodal for the forged gear blank considered here. There is a clear minimum near \(x = 0.35\). The function increases for both smaller and larger values. For \(x < 0.35\), the billet is too flat, and the material must undergo excessive radial flow to fill the rim, which increases damage. For \(x > 0.35\), the billet is too tall, and the upsetting stage produces excessive redundant shear, which also increases damage. The optimizer successfully locates the minimum.
I also compare the damage distribution for three different \(H_0/D_0\) values: 0.20, 0.35, and 1.00. The following table reports the maximum damage in the net forged gear blank and the volume error for each case. The case \(H_0/D_0 = 0.35\) has the lowest maximum damage and the smallest volume error. The case \(H_0/D_0 = 0.20\) has a slightly higher damage and a larger volume error because the billet is too flat. The case \(H_0/D_0 = 1.00\) has a much higher damage because the billet is too tall. This comparison confirms the optimality of the selected design.
| \(H_0/D_0\) | Maximum damage \(D_{\max}\) | Volume error \(g\) | Die filling |
|---|---|---|---|
| 0.20 | 0.41 | 0.012 | Marginal |
| 0.35 | 0.35 | 0.002 | Complete |
| 1.00 | 0.84 | 0.007 | Acceptable |
The choice of damage model is an important modeling decision. I use the Cockroft-Latham model because it is simple, computationally efficient, and widely used for forging. Other models, such as the Oyane model, the Rice-Tracey model, or the Gurson-Tvergaard-Needleman model, could also be used. Each model has its own calibration procedure and its own sensitivity to stress triaxiality. The optimization framework I present is not tied to a particular damage model. If a different model is used, the objective function and the finite element post-processing change, but the reduced gradient algorithm remains the same. This flexibility is an advantage of treating the finite element solver as a black-box function evaluator.
For the forged gear blank, the critical damage value \(C_{\text{cr}}\) is determined from a uniaxial tensile test or a compression test with controlled fracture. I assume \(C_{\text{cr}} = 1.0\) for the purpose of illustration. The maximum damage in the optimized forged gear blank is 0.35, which is well below the critical value. The safety margin is therefore about 65 percent. This margin is desirable because the actual production process may involve variations in temperature, friction, and material properties. A design with a large safety margin is more likely to produce a sound forged gear blank consistently.
I also consider the forming load. Although the objective is damage minimization, the forming load must not exceed the capacity of the press. The maximum load in the optimized design is slightly lower than in the initial design because the billet geometry is more favorable. This is a secondary benefit. The following table reports the maximum forming load for the initial and optimized designs. The load is normalized by the initial load. The optimized design requires about 92 percent of the initial load. This reduction is not the primary objective, but it is a welcome improvement.
| Quantity | Initial design | Optimized design |
|---|---|---|
| Normalized maximum forming load | 1.00 | 0.92 |
| Maximum damage | 0.94 | 0.35 |
| Volume error | 0.008 | 0.002 |
The optimization also affects the strain distribution. In the initial design, the effective strain is highly non-uniform. Some regions experience very large strain while others remain relatively undeformed. In the optimized design, the strain distribution is more uniform. This uniformity is beneficial because it reduces the risk of localized damage and improves the mechanical properties of the forged gear blank. A more uniform strain distribution also means that the subsequent heat treatment and machining operations are more predictable.
I can quantify the strain uniformity using the standard deviation of the effective strain in the net elements. Let \(\bar{\varepsilon}_i\) be the effective strain of element \(i\), and let \(N\) be the number of net elements. The mean effective strain is
$$ \bar{\varepsilon}_{\text{mean}} = \frac{1}{N} \sum_{i=1}^{N} \bar{\varepsilon}_i $$
and the standard deviation is
$$ s_{\varepsilon} = \sqrt{\frac{1}{N-1} \sum_{i=1}^{N} \left( \bar{\varepsilon}_i – \bar{\varepsilon}_{\text{mean}} \right)^2 } $$
The coefficient of variation is \(s_{\varepsilon} / \bar{\varepsilon}_{\text{mean}}\). For the initial design, the coefficient of variation is relatively high. For the optimized design, it is lower. This confirms that the optimized forged gear blank has a more uniform deformation history.
The following table summarizes the strain statistics for the initial and optimized designs. The improvement in uniformity is evident.
| Statistic | Initial design | Optimized design |
|---|---|---|
| Mean effective strain | 1.42 | 1.18 |
| Standard deviation | 0.68 | 0.41 |
| Coefficient of variation | 0.48 | 0.35 |
| Maximum effective strain | 3.21 | 2.24 |
The optimization also reduces the maximum effective strain in the forged gear blank. This is important because high effective strain can lead to excessive grain growth, flow localization, and other microstructural defects. By reducing the maximum strain, the optimized design improves the microstructural quality of the forged gear blank. The combination of lower damage and lower maximum strain is a strong indication that the optimized design is superior.
I now consider the practical implementation of the optimization. In a production environment, the initial billet geometry is often determined by the available bar stock. If the optimal \(H_0/D_0\) is 0.35, the billet must be cut to a specific length and diameter. This may require a change in the bar stock diameter or a change in the shearing process. The cost of changing the billet geometry is usually small compared with the cost of scrap or rework caused by a poor forged gear blank. Therefore, the optimization is economically justified. The reduced gradient method can be run offline, using a validated finite element model, and the resulting optimal design can be implemented in production without modifying the die cavity.
I also note that the optimization can be extended to multiple design variables. For example, the die temperature, the friction coefficient, and the ram speed could be included as additional variables. The reduced gradient method can handle multiple variables, provided that the gradient is computed with respect to each variable. The finite-difference approximation becomes more expensive as the number of variables increases, because each perturbation requires an additional simulation. For a small number of variables, the method remains practical. For a large number of variables, adjoint methods or surrogate models could be used. In this work, I focus on a single variable because the initial \(H_0/D_0\) is the most influential parameter for the forged gear blank studied here.
The following table lists the key results of the optimization in a concise form. I include the design variable, the objective function, the constraint, and the number of iterations. The table can be used as a summary of the optimization campaign.
| Item | Value |
|---|---|
| Design variable | Initial height-to-diameter ratio \(H_0/D_0\) |
| Initial design | 1.30 |
| Optimal design | 0.35 |
| Initial maximum damage | 0.94 |
| Optimal maximum damage | 0.35 |
| Damage reduction | 62.8% |
| Constraint tolerance | \(g \le 0.01\) |
| Optimal volume error | 0.002 |
| Number of iterations | 21 |
| Optimization algorithm | Reduced gradient with finite-difference sensitivity |
I can also express the optimization result in terms of the damage safety factor. Let the critical damage be \(C_{\text{cr}}\). The safety factor against fracture is
$$ \text{SF} = \frac{C_{\text{cr}}}{D_{\max}} $$
For the initial design, if \(C_{\text{cr}} = 1.0\), then \(\text{SF} = 1.0/0.94 \approx 1.06\). This is a very small safety margin. For the optimized design, \(\text{SF} = 1.0/0.35 \approx 2.86\). The safety factor improves by a factor of about 2.7. This is a substantial improvement. In engineering practice, a safety factor of 2.86 is much more acceptable than a safety factor of 1.06. The optimized forged gear blank is therefore much less likely to crack during forging or to fail in service.
I also consider the effect of the damage model on the optimal design. If I use a different damage model, the numerical value of the maximum damage will change, but the general trend is likely to remain the same. The optimal \(H_0/D_0\) may shift slightly, but it should remain in the range of 0.30 to 0.40. This robustness of the optimal design to the choice of damage model is reassuring. It suggests that the optimization is capturing a fundamental physical effect rather than a model-specific artifact. I recommend that the damage model be calibrated carefully before the optimization is used for production. The calibration should use experimental data for the same material and similar stress states as those encountered in the forged gear blank.
The finite element mesh sensitivity is another important consideration. I perform a mesh refinement study to ensure that the damage values are not significantly affected by the mesh density. The following table reports the maximum damage in the net forged gear blank for three mesh densities: coarse, medium, and fine. The medium mesh is used for the optimization. The fine mesh produces a maximum damage that is within 3 percent of the medium mesh. This indicates that the medium mesh is adequate for optimization purposes. The coarse mesh underestimates the maximum damage because it cannot resolve the strain gradients near the corners. I therefore avoid the coarse mesh.
| Mesh density | Number of elements | Maximum damage | Relative error |
|---|---|---|---|
| Coarse | 2,500 | 0.31 | 11.4% |
| Medium | 6,000 | 0.35 | 0.0% |
| Fine | 12,000 | 0.36 | 2.9% |
The computational time for the medium mesh is acceptable. The fine mesh requires roughly four times the computational time of the medium mesh. For the optimization loop, the medium mesh provides a good balance between accuracy and efficiency. If the optimization result is intended for final validation, I recommend rerunning the optimal design with the fine mesh to confirm the maximum damage. In this study, the fine mesh confirms that the optimal \(H_0/D_0\) remains 0.35 and that the maximum damage is still well below the critical value.
I also examine the influence of the friction coefficient on the optimal design. I rerun the optimization with friction coefficients of 0.2, 0.3, and 0.4. The following table reports the optimal \(H_0/D_0\) and the corresponding maximum damage for each friction coefficient. The optimal \(H_0/D_0\) changes only slightly, from 0.33 to 0.37. The maximum damage increases with increasing friction, as expected, because higher friction increases redundant deformation near the die surface. However, the optimal design remains in the same general range. This indicates that the optimization result is not overly sensitive to the friction coefficient. If the friction coefficient is uncertain, the optimal design can still be used as a good starting point.
| Friction coefficient | Optimal \(H_0/D_0\) | Maximum damage |
|---|---|---|
| 0.2 | 0.33 | 0.31 |
| 0.3 | 0.35 | 0.35 |
| 0.4 | 0.37 | 0.41 |
Similarly, I examine the influence of the billet temperature. I rerun the optimization with initial billet temperatures of 950 °C, 1000 °C, and 1050 °C. The optimal \(H_0/D_0\) remains close to 0.35. The maximum damage decreases as the temperature increases, because the flow stress decreases and the material flows more easily. This is consistent with the physics of hot forging. The optimal design is therefore robust with respect to reasonable variations in billet temperature.
| Billet temperature (°C) | Optimal \(H_0/D_0\) | Maximum damage |
|---|---|---|
| 950 | 0.36 | 0.39 |
| 1000 | 0.35 | 0.35 |
| 1050 | 0.34 | 0.31 |
The optimization framework can also be used to study the effect of die temperature. A higher die temperature reduces the heat loss from the billet and can improve material flow. However, it also increases die wear. The optimization can find a compromise if die temperature is included as a design variable. In this work, I keep the die temperature fixed at 300 °C. The optimal \(H_0/D_0\) is not strongly affected by the die temperature within the range of 250 °C to 350 °C. This is because the die temperature mainly affects the surface layers, while the maximum damage occurs in the interior or near the rim where the stress state dominates.
I also consider the possibility of preforming. In some forging processes, a preforming operation is used to redistribute material before the final impression. The optimization method described here could be extended to optimize the preform shape as well as the initial billet shape. This would involve multiple design variables and a more complex finite element model. The reduced gradient method could still be used, but the computational cost would increase. For the forged gear blank studied here, the single-stage process is sufficient, and the optimal initial billet geometry already produces a high-quality forged gear blank. A preforming operation is therefore not necessary.
The optimization results have practical implications for die design. The die cavity itself is not changed by the optimization. The same die can be used with the optimized billet. This is a major advantage because die manufacturing is expensive. The optimization only requires a change in the billet preparation. In many forging shops, the billet is cut from bar stock. Changing the billet length and diameter is relatively easy. The optimized forged gear blank can therefore be produced without modifying the die, which reduces the implementation cost and shortens the lead time.
I also note that the optimization can be performed before the die is manufactured. In the design stage, the die geometry is known from the forged gear blank drawing. The finite element model can be built and the optimization can be run. The result can then be used to specify the billet dimensions. This is a proactive approach to quality improvement. Instead of waiting for production trials and then correcting problems, the forging process can be optimized virtually. This reduces trial-and-error, shortens the development cycle, and lowers the cost of process development.
The following table summarizes the practical benefits of the optimized forged gear blank. The benefits are categorized as quality, cost, and productivity. The optimized design improves all three categories.
| Category | Benefit |
|---|---|
| Quality | Lower maximum damage, lower cracking risk, more uniform strain, better fatigue life |
| Cost | Less scrap, less rework, no die modification, lower forming load |
| Productivity | Fewer trial runs, faster process development, more stable production |
I also consider the limitations of the approach. The finite element model is an approximation of reality. The material properties, friction conditions, and heat transfer coefficients are uncertain. The optimization result is only as good as the model. I therefore recommend that the optimized design be validated with a small number of production trials before full-scale implementation. The production trials should measure the actual die filling, the forming load, and the occurrence of any surface or internal defects. If the trials confirm the simulation results, the optimized design can be adopted. If not, the model should be calibrated and the optimization rerun. This iterative approach is standard in virtual process design.
Another limitation is that the objective function uses the maximum damage in the net forging. This is a conservative criterion. In reality, fracture may not occur even if the maximum damage exceeds the critical value at a single point, because the point may be surrounded by material that constrains it. Conversely, fracture may occur at a point with a lower damage value if the stress state is more severe. The Cockroft-Latham model is a phenomenological model, and it cannot capture all the details of void growth and coalescence. Despite these limitations, the maximum damage criterion is widely used and provides a useful relative measure. It is particularly effective for comparing different designs, which is the purpose of this optimization.
I also acknowledge that the reduced gradient method is a local optimization method. It finds a local minimum, not necessarily the global minimum. In this one-dimensional problem, the function appears to be unimodal, so the local minimum is likely the global minimum. For higher-dimensional problems, a global search method such as a genetic algorithm or a particle swarm optimization could be used to explore the design space more thoroughly. However, global methods require many more function evaluations. The reduced gradient method is a good compromise for the forged gear blank problem because it converges quickly and produces a significant improvement.
The optimization framework is general and can be applied to other forged gear blanks and other forging processes. The key steps are always the same: define the objective function, choose the design variables, impose the constraints, build the finite element model, and run the optimization. The damage model can be selected based on the material and the process. The reduced gradient method can be replaced by another optimization algorithm if desired. The finite element solver can be replaced by another solver. The overall approach is modular and flexible.
I conclude that the initial height-to-diameter ratio of the billet is a critical parameter for the quality of a forged gear blank. By minimizing the maximum damage in the net forging, I obtain an optimal \(H_0/D_0\) of 0.35. This design reduces the maximum damage by about 62.8 percent compared with the initial design and ensures complete die filling. The optimized forged gear blank has a lower risk of cracking, a more uniform strain distribution, and a higher safety factor against fracture. The optimization is performed with a finite element solver and a reduced gradient algorithm. The method is efficient, robust, and practical for industrial use. The results demonstrate that damage-minimized process design can significantly improve the quality of a forged gear blank without changing the die cavity.
For future work, I would extend the optimization to include additional design variables such as the billet temperature, the die temperature, the friction coefficient, and the ram speed. I would also investigate the use of adaptive remeshing and higher-order elements to improve the accuracy of the damage prediction. I would validate the optimized design with production trials and measure the actual fatigue life of the forged gear blank. Finally, I would explore the use of surrogate models to reduce the computational cost of the optimization, making it feasible to optimize more complex forged gear blanks with many design variables. These extensions would further strengthen the link between numerical optimization and industrial forging practice.