In my work on precision investment casting for aerospace titanium alloy structural components, I treat wax pattern injection forming as a decisive stage for dimensional accuracy, surface integrity, and final casting quality. The investment casting chain begins with a wax pattern whose geometry is later replicated by a ceramic shell and finally by molten titanium alloy. Any error in the wax pattern, especially an error caused by non-uniform filling, cooling, shrinkage, or warpage, can propagate through the entire investment casting route. Because numerical simulation is one of the most effective tools for optimizing wax injection, the reliability of the simulation depends strongly on the viscosity model of the wax. I therefore focused on a high-precision parameter modeling method for the Cross-WLF viscosity model of a wax material used in investment casting.
The material I selected is a wax designated Kc-4017b, which is applied in investment casting for titanium alloy components. Its density is 0.95 g/cm3, its drop point is 94 degrees C, and its softening point is 79 degrees C. The viscosity behavior shows clear shear-thinning characteristics, meaning that the apparent viscosity decreases rapidly as shear rate increases and then tends toward a more stable value. Temperature also has a strong influence: at the same shear rate, a higher temperature produces a lower viscosity. This coupled dependence on shear rate and temperature makes a simple constant-viscosity assumption inadequate for investment casting simulation. I therefore adopted the Cross-WLF model, which combines a Cross shear-thinning term with a Williams-Landel-Ferry temperature-sensitive term.
My central objective is to solve the high-dimensional nonlinear parameter fitting problem of the Cross-WLF model with high accuracy and numerical stability. In investment casting, wax injection involves complex flow, rapid cooling near the mold wall, packing, and subsequent warpage. A poor viscosity model can shift the predicted melt front, pressure field, temperature field, shrinkage distribution, and warpage pattern. Since the commercial material database of Moldflow does not contain the required Cross-WLF parameters for this wax, I needed to obtain the seven parameters from experimental viscosity data. I propose a quasi-Newton method that integrates automatic differentiation and singular value decomposition regularization. I validated the method through annular stepped wax pattern simulation and actual injection experiments with blue-light scanning.
Material Behavior and Rheological Data
I measured the viscosity of the wax at 60 degrees C, 62 degrees C, and 64 degrees C over a range of shear rates. The experimental curves show a typical non-Newtonian response. At low shear rate, the viscosity is high because molecular entanglements and intermolecular interactions resist flow. As shear rate increases, the microstructure is progressively oriented and disentangled, so viscosity decreases. At sufficiently high shear rate, the viscosity approaches a lower plateau. The temperature increase lowers the zero-shear viscosity and reduces the resistance to flow, which is consistent with increased molecular mobility.
| Property | Value | Unit |
|---|---|---|
| Wax material | Kc-4017b | – |
| Density | 0.95 | g/cm3 |
| Drop point | 94 | degrees C |
| Softening point | 79 | degrees C |
| Rheological class | Non-Newtonian, shear-thinning | – |
| Testing temperatures | 60, 62, 64 | degrees C |
The measured data indicate that both shear rate and temperature must appear in the viscosity model. For investment casting, this matters because the wax experiences a wide range of shear rates as it passes through the gate, runner, and cavity, while simultaneously cooling against the die. A model that captures only shear rate or only temperature cannot describe the coupled physical behavior. The Cross-WLF model is therefore the most suitable choice for my investment casting simulation.
Cross-WLF Viscosity Model
The Cross-WLF model describes apparent viscosity as a function of shear rate, temperature, and pressure. I write the general form as follows:
$$
\eta(\dot{\gamma},T,p)=\frac{\eta_0(T,p)}{1+\left(\frac{\eta_0(T,p)\dot{\gamma}}{\tau^*}\right)^{1-n}}
$$
where \(\eta\) is the apparent viscosity in Pa·s, \(\eta_0\) is the zero-shear viscosity in Pa·s, \(\dot{\gamma}\) is the shear rate in s-1, \(\tau^*\) is the characteristic relaxation stress in Pa, and \(n\) is the non-Newtonian power-law index. When \(n<1\), the model predicts shear thinning. A smaller \(n\) corresponds to stronger shear-thinning behavior.
The temperature dependence of the zero-shear viscosity is represented by the WLF equation:
$$
\eta_0(T,p)=D_1 \exp\left[-\frac{A_1(T-T^*)}{A_2+(T-T^*)}\right]
$$
The reference temperature \(T^*\) is written as:
$$
T^*=D_2+D_3 p
$$
In these equations, \(T\) is the absolute temperature in K, \(p\) is pressure in Pa, \(D_1\) is the zero-shear viscosity coefficient at the reference temperature in Pa·s, \(D_2\) is the glass transition temperature at low pressure in K, \(D_3\) is the pressure coefficient, \(A_1\) is the WLF temperature sensitivity parameter, and \(A_2\) is the WLF temperature shift parameter. For small-molecule wax systems in investment casting, I take \(D_3=0\), because the pressure dependence of the glass transition temperature is negligible compared with the strong temperature and shear-rate effects.
| Symbol | Meaning | Unit |
|---|---|---|
| \(\eta\) | Apparent viscosity | Pa·s |
| \(\eta_0\) | Zero-shear viscosity | Pa·s |
| \(\dot{\gamma}\) | Shear rate | s-1 |
| \(\tau^*\) | Characteristic relaxation stress | Pa |
| \(n\) | Non-Newtonian index | dimensionless |
| \(T\) | Absolute temperature | K |
| \(T^*\) | Reference glass transition temperature | K |
| \(D_1\) | Reference zero-shear viscosity coefficient | Pa·s |
| \(D_2\) | Low-pressure glass transition temperature | K |
| \(D_3\) | Pressure coefficient | K/Pa |
| \(A_1\) | WLF temperature sensitivity parameter | dimensionless |
| \(A_2\) | WLF temperature shift parameter | K |
Parameter Identification Challenge
I define the parameter vector for the Cross-WLF model as:
$$
\theta=\left[n,\tau^*,D_1,D_2,D_3,A_1,A_2\right]^T
$$
Since \(D_3\) is fixed to zero for the wax, the effective optimization dimension is six. Even so, the model remains strongly nonlinear because \(\eta_0\) appears inside the Cross term and is itself an exponential function of temperature. The objective is to minimize the squared difference between experimental viscosity \(y_i\) and model prediction \(\hat{y}_i\):
$$
L(\theta)=\sum_{i=1}^{N}\left(y_i-\hat{y}_i(\theta)\right)^2
$$
A common approach is simple least squares. However, in my experience with investment casting wax, simple least squares relies mainly on first-order gradient information and can become trapped in local minima. It is also sensitive to initial guesses. A Newton method can use second-order information, but it requires the Hessian matrix and its inverse. The computational complexity of inverting a dense Hessian is of order \(O(n^3)\), and the Hessian may be singular or ill-conditioned. If the condition number \(\kappa(H)\) exceeds \(10^6\), the Newton step can become unstable or divergent. This is particularly problematic for wax viscosity data, where viscosity changes steeply with temperature and shear rate.
| Method | Information Used | Main Limitation | Typical Complexity |
|---|---|---|---|
| Simple least squares | First-order gradient or linearized sensitivity | Local minima, poor nonlinear coupling, initial-value sensitivity | Low per iteration but may need many trials |
| Newton method | Gradient and full Hessian | High complexity, Hessian inversion, numerical instability | \(O(n^3)\) |
| Quasi-Newton method with BFGS | Gradient and approximate inverse Hessian | Requires careful line search and regularization | \(O(n^2)\) |
| Proposed AD-SVD quasi-Newton method | Automatic differentiation, BFGS update, SVD regularization | More implementation effort but robust for investment casting wax | \(O(n^2)\) |
Quasi-Newton Method with Automatic Differentiation and SVD Regularization
I formulate the update rule of the quasi-Newton method as:
$$
\theta_{k+1}=\theta_k-\alpha_k H_k^{-1}\nabla f(\theta_k)
$$
Here \(k\) is the iteration index, \(\alpha_k\) is the step size, \(H_k^{-1}\) is an approximation to the inverse Hessian matrix, and \(\nabla f(\theta_k)\) is the gradient of the objective function. Instead of computing the Hessian directly, I update \(H_k^{-1}\) by the BFGS formula:
$$
H_{k+1}^{-1}=\left(I-\rho_k s_k y_k^T\right)H_k^{-1}\left(I-\rho_k y_k s_k^T\right)+\rho_k s_k s_k^T
$$
where:
$$
s_k=\theta_{k+1}-\theta_k
$$
$$
y_k=\nabla f(\theta_{k+1})-\nabla f(\theta_k)
$$
$$
\rho_k=\frac{1}{y_k^T s_k}
$$
The BFGS update preserves positive definiteness under suitable conditions, which is important for stable descent. I use automatic differentiation to compute gradients and Jacobian-vector products accurately. For a sum-of-squares objective, the gradient can be written as:
$$
\nabla_\theta L(\theta)=2\sum_{i=1}^{N}\left(\hat{y}_i(\theta)-y_i\right)\nabla_\theta \hat{y}_i(\theta)
$$
Automatic differentiation avoids finite-difference truncation error and provides exact derivatives up to machine precision. This is valuable in investment casting wax modeling because the Cross-WLF equation contains exponential and power-law terms that can amplify small derivative errors.
I also use the Armijo condition to select the step size. The condition is:
$$
f\left(\theta_k-\alpha p_k\right)\le f(\theta_k)+c_1\alpha \nabla f(\theta_k)^T p_k
$$
where \(p_k=-H_k^{-1}\nabla f(\theta_k)\) is the search direction and \(c_1\) is a small positive constant, typically \(10^{-4}\). This line search ensures sufficient decrease of the objective and improves global convergence behavior.
When the approximate Hessian matrix \(A\) is ill-conditioned or nearly singular, I apply singular value decomposition:
$$
A=U\Sigma V^T
$$
The inverse is computed through regularized singular values:
$$
A^{-1}=V\Sigma^+ U^T
$$
If a singular value \(\sigma_i\) is smaller than a threshold \(\epsilon\sigma_{\max}\), I truncate or regularize it. This prevents numerical explosion and stabilizes the quasi-Newton iteration. The combination of automatic differentiation, BFGS updating, Armijo line search, and SVD regularization forms the core of my proposed method for high-precision viscosity modeling in investment casting.
| Step | Operation | Purpose |
|---|---|---|
| 1 | Initialize \(\theta_0\) from physical bounds | Keep parameters in a meaningful range |
| 2 | Evaluate \(L(\theta_k)\) and \(\nabla L(\theta_k)\) by automatic differentiation | Obtain exact gradient information |
| 3 | Compute search direction \(p_k=-H_k^{-1}\nabla L(\theta_k)\) | Generate a descent direction |
| 4 | Apply Armijo line search for \(\alpha_k\) | Ensure sufficient decrease |
| 5 | Update \(\theta_{k+1}=\theta_k+\alpha_k p_k\) | Improve parameter estimate |
| 6 | Update \(H_{k+1}^{-1}\) by BFGS | Avoid full Hessian inversion |
| 7 | Apply SVD regularization if \(\kappa(A)>10^6\) | Control ill-conditioning |
| 8 | Check convergence of \(L(\theta)\) and \(\nabla L(\theta)\) | Stop when tolerance is reached |
Initial Bounds and Numerical Stabilization
I constrained the Cross-WLF parameters within physically meaningful ranges. These bounds prevent the optimizer from moving into non-physical regions where the model may predict negative viscosity or unrealistic temperature dependence. The bounds I used are summarized below.
| Parameter | Lower Bound | Upper Bound | Physical Role |
|---|---|---|---|
| \(n\) | 0.10 | 0.90 | Shear-thinning intensity |
| \(\tau^*\) | 0.01 | 10.00 | Relaxation stress transition |
| \(D_1\) | \(1.0\times 10^{10}\) | \(1.0\times 10^{20}\) | Reference zero-shear viscosity |
| \(D_2\) | 200 | 300 | Low-pressure glass transition temperature |
| \(D_3\) | 0 | 0 | Fixed for wax |
| \(A_1\) | 1 | 100 | Temperature sensitivity |
| \(A_2\) | 10 | 200 | Temperature shift |
These bounds are not arbitrary. They reflect the known thermal behavior of the wax and the physical meaning of the Cross-WLF model. For investment casting, a stable parameter set is more important than a mathematically perfect fit that violates physical constraints. The regularized quasi-Newton method allows me to obtain a stable optimum while respecting these bounds.
Numerical Simulation Setup for Investment Casting
I constructed an annular stepped wax pattern that represents the type of varying wall thickness found in aerospace titanium alloy structural components. The stepped geometry is useful because it creates non-uniform flow resistance, non-uniform cooling rates, and non-uniform shrinkage. These features are common in investment casting and are precisely where viscosity model errors become visible.
I imported the geometry into Moldflow and generated a double-layer mesh. The global edge length was set to 2 mm. I checked mesh quality metrics such as aspect ratio and element overlap to ensure that the mesh satisfied the accuracy requirements of the simulation. The process parameters matched the actual injection experiment, as shown in the table below.
| Parameter | Value | Unit |
|---|---|---|
| Wax temperature | 65 | degrees C |
| Mold temperature | 25 | degrees C |
| Packing pressure | 1 | MPa |
| Packing time | 60 | s |
| Gate diameter | 8 | mm |
| Wax-mold heat transfer coefficient | 2500 | W/(m2·K) |
| Mesh edge length | 2 | mm |
| Analysis sequence | Cooling-Filling-Packing-Warpage | – |
The analysis sequence covered the full investment casting wax injection process: cooling, filling, packing, and warpage. This is important because the viscosity model affects each stage. During filling, viscosity controls the melt front and pressure drop. During packing, it controls pressure transmission and compensation. During cooling, it affects the relaxation of molecular orientation and shrinkage. During warpage, the accumulated stress and shrinkage determine the final displacement field.
Injection Experiment and Blue-Light Scanning
I performed the actual injection experiment on an MPI-25 t wax injection machine. The wax temperature was 65 degrees C, the mold temperature was 25 degrees C, the packing pressure was 1 MPa, and the packing time was 60 s. The gate diameter was 8 mm. I repeated the experiment five times under the same conditions to ensure repeatability. After injection and cooling, I used blue-light scanning to record the displacement and dimensional changes of the wax pattern. The scanned data were then compared with the Moldflow simulation results.
This experimental validation is critical for investment casting because simulation accuracy cannot be judged by curve fitting alone. A viscosity model may fit rheological data well but still produce poor filling and warpage predictions. By combining Cross-WLF parameter fitting, Moldflow simulation, and blue-light scanning, I evaluated the proposed method from the material level to the component level.
Fitting Results and Comparison
I compared the proposed quasi-Newton method with simple least squares. The fitted Cross-WLF parameters are listed below.
| Parameter | Quasi-Newton Method | Least Squares Method | Unit |
|---|---|---|---|
| \(n\) | 0.62 | 0.72 | dimensionless |
| \(\tau^*\) | 0.61 | 0.11 | Pa |
| \(D_1\) | \(1.08\times 10^{17}\) | \(1.77\times 10^{17}\) | Pa·s |
| \(D_2\) | 263.85 | 256.66 | K |
| \(D_3\) | 0 | 0 | K/Pa |
| \(A_1\) | 55.85 | 78.13 | dimensionless |
| \(A_2\) | 51.38 | 99.98 | K |
The quasi-Newton method produced a lower power-law index, \(n=0.62\), compared with \(n=0.72\) from least squares. A lower \(n\) indicates stronger shear-thinning behavior, which is consistent with the measured viscosity data. The characteristic relaxation stress also differed significantly: \(0.61\) Pa for the quasi-Newton method and \(0.11\) Pa for least squares. The WLF parameters \(A_1\) and \(A_2\) also changed substantially, reflecting the sensitivity of the temperature term. These differences show that the optimization method has a strong effect on the final parameter set, even when the same experimental data are used.
I evaluated the fit quality using the coefficient of determination:
$$
R^2=1-\frac{\sum_{i=1}^{N}\left(y_i-\hat{y}_i\right)^2}{\sum_{i=1}^{N}\left(y_i-\bar{y}\right)^2}
$$
I also used the root-mean-square error:
$$
RMSE=\sqrt{\frac{1}{N}\sum_{i=1}^{N}\left(y_i-\hat{y}_i\right)^2}
$$
The comparison is summarized below.
| Metric | Quasi-Newton Method | Least Squares Method | Interpretation |
|---|---|---|---|
| \(R^2\) | 0.989 | 0.901 | Higher is better |
| Residual distribution | Uniform and centered near zero | Less uniform, larger local deviations | Quasi-Newton is more stable |
| High shear-rate behavior | Close to experimental data | Progressive deviation | Important for gate and thin sections |
| Temperature coupling | Accurate across 60-64 degrees C | Weaker at higher temperature | Important for investment casting cooling |
| Local minima risk | Low | High | Quasi-Newton is more robust |
At 60 degrees C, the wax molecules are more entangled, so viscosity is more sensitive to shear rate. The quasi-Newton method captured this sensitivity more accurately. At 62 degrees C and 64 degrees C, the viscosity decreased, but the quasi-Newton method still maintained higher accuracy. This cross-temperature stability is essential for investment casting because the wax passes through a temperature gradient during filling and cooling.
Residual Analysis
I analyzed the residuals in both surface and histogram forms. For the quasi-Newton method, the residual surface was flatter and the color variation was smoother, indicating that errors were distributed more evenly across shear rate and temperature. For least squares, the residual surface showed pronounced local peaks, meaning that certain combinations of shear rate and temperature produced larger errors. These peaks are dangerous for investment casting simulation because they can coincide with the actual flow conditions in the gate or thin walls.
The residual histograms provided additional evidence. The quasi-Newton residuals were concentrated near zero, while the least-squares residuals had a wider spread and a heavier tail. A heavy tail means that some data points are poorly represented, which can lead to local errors in the predicted viscosity field. In an investment casting simulation, even a local viscosity error can shift the melt front and change the pressure distribution.
| Residual Feature | Quasi-Newton Method | Least Squares Method |
|---|---|---|
| Surface smoothness | High | Low |
| Local peaks | Few | Several |
| Histogram spread | Narrow | Wide |
| Large-error frequency | Low | Higher |
| Suitability for investment casting simulation | High | Moderate |
Moldflow Simulation Results
I imported both parameter sets into Moldflow and simulated the injection process. The filling-stage temperature field showed that the melt front remained at approximately 62 degrees C, which matched the set temperature. Near the mold wall, the melt temperature dropped rapidly toward the mold temperature, and the viscosity increased sharply to a solid-like value. This behavior is expected in investment casting wax injection because the die acts as a strong heat sink.
At the melt front, the quasi-Newton parameter set predicted a viscosity of about 1.36 Pa·s, while the least-squares parameter set predicted about 0.86 Pa·s. The quasi-Newton value is more consistent with the measured viscosity data. This difference is not trivial. A lower viscosity prediction can cause the simulation to underestimate flow resistance, overestimate filling ease, and produce a different pressure history. In investment casting, such differences can affect the predicted location of weld lines, air traps, and shrinkage porosities.
| Simulation Quantity | Quasi-Newton Parameters | Least-Squares Parameters | Experimental Relevance |
|---|---|---|---|
| Melt front temperature | About 62 degrees C | About 62 degrees C | Matches set temperature |
| Melt front viscosity | 1.36 Pa·s | 0.86 Pa·s | Quasi-Newton agrees better with measured data |
| Wall viscosity | Solid-like high value | Solid-like high value | Consistent with rapid cooling |
| Warpage tendency | Slightly lower | Slightly higher | Related to cooling uniformity |
The cooling behavior also differed. The quasi-Newton parameters produced a slightly higher viscosity, which corresponds to more tightly entangled molecular chains, especially at lower temperature. During cooling, this slower relaxation and crystallization behavior makes the temperature field and shrinkage field evolve more synchronously. As a result, the temperature gradient across the stepped sections was smaller, and the predicted warpage was slightly lower than that obtained from the least-squares parameters. This is favorable for investment casting because lower warpage improves wax pattern dimensional accuracy and reduces shell cracking risk during subsequent processing.
Warpage and Shrinkage Validation
The annular stepped geometry created large differences in volumetric shrinkage between sections. The regions where shrinkage changed abruptly were the most likely to exceed dimensional tolerance. In the simulation with quasi-Newton parameters, the second stepped surface and the gate region showed clear transitions in average volumetric shrinkage. The blue-light scanning data showed displacement transitions at the same locations. The physical wax pattern also exhibited sink marks in those regions. This agreement confirms that the viscosity model obtained by the quasi-Newton method improves the reliability of investment casting simulation.
I compared the simulated displacement trend with the measured displacement trend. The overall trend matched well. The critical deformation regions identified by simulation were consistent with the experimental scan. This is important because the goal of simulation in investment casting is not only to predict global shrinkage but also to locate local defects. The proposed viscosity model provides a more reliable basis for that task.
| Validation Item | Simulation Result | Blue-Light Scan Result | Agreement |
|---|---|---|---|
| Second stepped surface | Shrinkage transition | Displacement transition | High |
| Gate region | Shrinkage transition | Displacement transition | High |
| Overall displacement trend | Predicted | Measured | High |
| Sink mark locations | Predicted | Observed | Consistent |
Discussion
The results show that the main difficulty in wax viscosity modeling for investment casting is not merely the choice of equation but the numerical method used to identify parameters. The Cross-WLF model is flexible enough to describe shear thinning and temperature sensitivity, but its parameters are correlated. A change in \(D_1\) can be partly compensated by a change in \(A_1\) or \(A_2\), and a change in \(n\) can interact with \(\tau^*\). This coupling creates a rugged objective landscape. Simple least squares follows the local gradient and may settle in a local valley. The quasi-Newton method uses gradient history and an approximate inverse Hessian, so it can follow curvature information and escape shallow local minima.
Automatic differentiation further improves the optimization because it computes derivatives without finite-difference noise. In a model containing exponential terms such as \(\exp[-A_1(T-T^*)/(A_2+T-T^*)]\), finite-difference gradients can be inaccurate if the step size is not carefully chosen. Automatic differentiation avoids this issue and provides consistent derivative information throughout the iteration. This is particularly useful when fitting data at multiple temperatures, as I did for the investment casting wax.
SVD regularization addresses the ill-conditioning that appears when the BFGS approximation becomes nearly singular. In the early iterations, the approximate Hessian may be a poor representation of the true curvature. If the inverse is taken without regularization, small singular values can produce enormous parameter updates. SVD truncation suppresses those directions and keeps the step within a physically meaningful range. This makes the method more robust for investment casting applications, where the material data may be limited and the model must remain stable.
The Moldflow simulation confirmed that the fitted parameters affect not only the viscosity curve but also the predicted filling and warpage. The melt front viscosity predicted by the quasi-Newton parameters was closer to the measured value. The temperature field and shrinkage field evolved more uniformly. The displacement trend matched the blue-light scan. These findings support the use of the proposed method as a practical tool for generating wax viscosity data for investment casting simulation.
From an industrial perspective, the ability to model wax viscosity accurately can reduce trial-and-error iterations in investment casting. A reliable Cross-WLF parameter set allows engineers to optimize injection temperature, mold temperature, packing pressure, packing time, and gate design before cutting a die. This shortens development cycles and lowers cost. For titanium alloy investment casting, where dimensional tolerance and surface quality are demanding, the benefit is significant.
Practical Implications for Investment Casting
I see several practical implications of this work for investment casting. First, the Cross-WLF parameters should be obtained with a method that captures nonlinear coupling. Second, the parameter set should be validated not only by \(R^2\) but also by Moldflow simulation and physical scanning. Third, the viscosity model should be checked at the temperatures and shear rates that actually occur in the gate, runner, and cavity. Fourth, residual analysis should be used to identify regions where the model may fail. Fifth, the simulation should include cooling, filling, packing, and warpage because viscosity affects all four stages.
| Practical Action | Reason | Expected Benefit in Investment Casting |
|---|---|---|
| Use nonlinear parameter fitting | Cross-WLF parameters are strongly coupled | More accurate viscosity field |
| Validate with component-level simulation | Curve fit alone is insufficient | Better prediction of filling and warpage |
| Apply SVD regularization | Hessian may be ill-conditioned | Stable and repeatable parameter identification |
| Use automatic differentiation | Finite differences can be noisy | Accurate gradients and faster convergence |
| Compare with blue-light scanning | Experimental displacement is ground truth | Confidence in investment casting simulation |
Mathematical Summary
For clarity, I summarize the complete Cross-WLF formulation and the optimization procedure below.
$$
\eta(\dot{\gamma},T,p)=\frac{\eta_0(T,p)}{1+\left(\frac{\eta_0(T,p)\dot{\gamma}}{\tau^*}\right)^{1-n}}
$$
$$
\eta_0(T,p)=D_1 \exp\left[-\frac{A_1(T-T^*)}{A_2+(T-T^*)}\right]
$$
$$
T^*=D_2+D_3 p
$$
$$
L(\theta)=\sum_{i=1}^{N}\left(y_i-\hat{y}_i(\theta)\right)^2
$$
$$
\theta_{k+1}=\theta_k-\alpha_k H_k^{-1}\nabla L(\theta_k)
$$
$$
H_{k+1}^{-1}=\left(I-\rho_k s_k y_k^T\right)H_k^{-1}\left(I-\rho_k y_k s_k^T\right)+\rho_k s_k s_k^T
$$
$$
\rho_k=\frac{1}{y_k^T s_k}
$$
$$
f\left(\theta_k-\alpha p_k\right)\le f(\theta_k)+c_1\alpha \nabla f(\theta_k)^T p_k
$$
$$
A=U\Sigma V^T
$$
$$
A^{-1}=V\Sigma^+ U^T
$$
These equations define the model and the numerical strategy I used to obtain high-precision viscosity parameters for investment casting wax injection. The combination of physical constraints, automatic differentiation, BFGS updating, Armijo line search, and SVD regularization is the key to the improved performance.
Comparison of Methods and Error Metrics
I summarize the main quantitative comparison between the proposed quasi-Newton method and the traditional least-squares method in the following table. The quasi-Newton method achieved a higher coefficient of determination, a more uniform residual distribution, and a more physically meaningful viscosity prediction at the melt front. These improvements are directly relevant to investment casting because they reduce uncertainty in filling, packing, cooling, and warpage simulation.
| Comparison Item | Quasi-Newton Method | Least Squares Method | Meaning for Investment Casting |
|---|---|---|---|
| Coefficient of determination | 0.989 | 0.901 | Higher confidence in viscosity data |
| Residual distribution | Narrow and centered | Wide and scattered | Fewer local prediction errors |
| High shear-rate fit | Accurate | Deviates | Better gate and thin-wall prediction |
| Temperature coupling | Stable from 60 to 64 degrees C | Less stable | Better cooling and shrinkage prediction |
| Melt front viscosity | 1.36 Pa·s | 0.86 Pa·s | Quasi-Newton is closer to measurement |
| Warpage prediction | Closer to blue-light scan | Less consistent | Improved dimensional control |
Why Investment Casting Benefits from This Modeling Approach
Investment casting is a chain of precision replication steps. The wax pattern is the first replicated geometry, and its accuracy influences every downstream step. If the wax pattern warps, the ceramic shell may distort, the mold cavity may shift, and the final titanium alloy casting may exceed tolerance. Because wax injection is a non-Newtonian, non-isothermal, transient process, accurate viscosity modeling is essential. My proposed method provides a reliable way to obtain Cross-WLF parameters for wax materials that are not available in standard simulation databases.
The method is especially useful for aerospace investment casting, where components often have thin walls, thick bosses, abrupt thickness changes, and complex annular features. In these geometries, the shear rate and cooling rate vary strongly from point to point. A viscosity model that is accurate only in a narrow range is not sufficient. The quasi-Newton method with automatic differentiation and SVD regularization provides stable fitting across a wider range of temperatures and shear rates, which translates into better predictions for the entire investment casting process.
Limitations and Future Work
I acknowledge that the present study focuses on one wax material and one representative annular stepped geometry. Different wax blends may have different molecular structures, filler contents, and thermal histories. The same numerical framework can be applied, but the parameter bounds and initial guesses may need adjustment. Future work could extend the method to multiple wax materials, pressure-dependent viscosity, crystallization effects, and transient rheology. It would also be useful to couple the viscosity model with thermal contact resistance and mold deformation, because those factors also influence wax pattern accuracy in investment casting.
Another direction is to reduce the experimental burden by combining sparse rheological measurements with physics-informed constraints. Automatic differentiation can naturally incorporate such constraints into the loss function. For investment casting, a physics-informed viscosity model could enforce monotonic shear thinning, positive viscosity, and physically reasonable WLF parameters. This would improve robustness when data are limited.
Conclusion
I developed a high-precision modeling method for the Cross-WLF viscosity parameters of a wax used in investment casting. The method combines a quasi-Newton framework, automatic differentiation, BFGS inverse Hessian updates, Armijo line search, and SVD regularization. The main conclusions are as follows.
First, the quasi-Newton method achieved \(R^2=0.989\), which is significantly higher than the value of \(0.901\) obtained by simple least squares. The residual distribution was more uniform, and the model captured the coupled effects of shear rate and temperature more accurately.
Second, the fitted Cross-WLF parameters produced reliable Moldflow predictions for an annular stepped wax pattern. The predicted melt front viscosity was closer to the measured viscosity, and the temperature and shrinkage fields evolved more consistently.
Third, the simulation results agreed with actual injection experiments and blue-light scanning. The critical deformation regions, including the stepped surface and gate region, matched the measured displacement transitions. Sink marks observed on the physical wax pattern also appeared in the simulated risk regions.
Fourth, the proposed method solves the high-dimensional nonlinear parameter identification problem for wax viscosity in investment casting. It fills the gap in standard simulation databases for wax-based materials and provides reliable parameter support for numerical simulation of wax pattern injection forming.
Overall, my work shows that accurate wax viscosity modeling is a key enabling step for precision investment casting of titanium alloy structural components. By improving the Cross-WLF parameter identification, the investment casting simulation becomes more reliable, the wax pattern dimensional accuracy can be better controlled, and the development cycle for complex aerospace components can be shortened.
| Final Outcome | Result | Impact on Investment Casting |
|---|---|---|
| Best fitting method | Quasi-Newton with AD and SVD | High-precision wax viscosity parameters |
| \(R^2\) | 0.989 | Improved confidence in simulation |
| Moldflow validation | Consistent with injection experiment | Reliable filling and warpage prediction |
| Blue-light scanning agreement | Matched displacement trend | Better dimensional control of wax patterns |
| Application | Aerospace titanium alloy investment casting | Shorter development cycle and lower cost |
In my view, this modeling approach can serve as a practical foundation for future investment casting simulation workflows. It provides a route from experimental viscosity data to stable Cross-WLF parameters, and from those parameters to reliable predictions of wax injection forming. For complex titanium alloy investment casting, such a route is essential for achieving high dimensional accuracy and robust process control.