The manufacturing of large, thin-walled components for aerospace applications, such as cabin and pressure vessel segments, presents significant challenges. These aerospace castings often feature complex geometries with uneven wall thickness and integral stiffening ribs. This non-uniformity disrupts ideal solidification patterns, making feeding difficult and leading to the formation of shrinkage porosity and micro-shrinkage defects, which can compromise the structural integrity and pressure tightness of the final component. Low-pressure die casting (LPDC) is a prevalent technique for producing such high-integrity aerospace castings. In LPDC, molten metal is forced upward into the die cavity by applying a controlled gas pressure, resulting in a tranquil, non-turbulent fill. The sustained pressure during solidification enhances feeding, potentially eliminating the need for extensive risering systems. However, optimizing the numerous interacting process parameters—such as pouring temperature, mold temperature, filling time, and holding pressure—to eliminate defects remains a complex, experience-dependent task often requiring numerous trial runs.
Traditionally, numerical simulation based on physical mechanism models (e.g., solving Navier-Stokes and energy conservation equations) has been the primary tool for virtual process optimization. While effective, high-fidelity simulations of a complex aerospace casting are computationally expensive, with a single run potentially taking several hours. This cost prohibits exhaustive exploration of the vast parameter space. The emergence of data-driven methods, particularly machine learning (ML), offers a powerful complementary approach. By building surrogate models from simulation or experimental data, ML can predict casting outcomes with millisecond speed, enabling rapid design space exploration and multi-objective optimization. This work details a methodology that synergistically combines physics-based numerical simulation with a Gaussian Process Regression (GPR) surrogate model and a Genetic Algorithm (GA) to efficiently design and optimize the LPDC process for a representative thin-walled cylindrical aerospace casting.
Methodology: Integrating Simulation and Data-Driven Learning
The core of the methodology involves three interconnected stages: 1) Generating a high-quality dataset using high-fidelity numerical simulation, 2) Training a fast and accurate machine learning surrogate model on this data, and 3) Employing an evolutionary algorithm to find the optimal process parameters guided by the surrogate.
1. Numerical Simulation as the Data Source
The subject is a cylindrical Al-Si alloy (similar to A356/357) casting with a height of 730 mm, an outer diameter of 380 mm, and an average wall thickness of 10 mm, featuring internal reinforcement ribs. A bottom-filling gating system with multiple vertical gates was designed to promote directional solidification. The key thermal-physical properties of the alloy, crucial for accurate simulation, are summarized below.
| Property | Variation with Temperature |
|---|---|
| Thermal Conductivity (k) | Decreases slightly in the mushy zone. |
| Density (ρ) | Increases upon solidification. |
| Specific Heat (Cp) | Shows a peak corresponding to latent heat release. |
| Solid Fraction (fs) | Defined by the alloy’s phase diagram. |
The simulation solves the governing equations for fluid flow, heat transfer, and solidification, incorporating the Volume of Fluid (VOF) method for free surface tracking during mold filling. The primary output metric for defect prediction is the volume of shrinkage porosity. This is predicted using the widely adopted Niyama criterion, which relates the local thermal conditions to the likelihood of pore formation. The criterion states that a region is prone to shrinkage porosity if the Niyama value (Ny) falls below a critical threshold. It is calculated as:
$$
Ny = \frac{G}{\sqrt{\dot{T}}}
$$
where \(G\) is the temperature gradient (°C/cm) and \(\dot{T}\) is the cooling rate (°C/s). A lower Ny value indicates a higher risk of microporosity due to difficult interdendritic feeding.
The process parameters considered for optimization, along with their investigated ranges, are listed in the following table. A full factorial design across four parameters at three levels each would require \(3^4 = 81\) simulation runs.
| Process Parameter | Symbol | Lower Level | Middle Level | Upper Level |
|---|---|---|---|---|
| Pouring Temperature | Tp (°C) | 700 | 740 | 780 |
| Mold Temperature | Tm (°C) | 250 | 275 | 300 |
| Filling Time | tf (s) | 8 | 10 | 12 |
| Holding Pressure | Ph (kPa) | 180 | 200 | 220 |
2. Building the Machine Learning Surrogate Model
To avoid the prohibitive cost of 81 full simulations, an initial dataset is constructed using a classic L9(3^4) orthogonal array design, which requires only 9 strategically chosen simulations. The results from these 9 runs form the training dataset \(\mathcal{D}\) for the machine learning model:
$$
\mathcal{D} = \{(\mathbf{x_i}, y_i)\}_{i=1}^{9}, \quad \text{where} \quad \mathbf{x_i} = [T_p^{(i)}, T_m^{(i)}, t_f^{(i)}, P_h^{(i)}] \quad \text{and} \quad y_i = V_{porosity}^{(i)}
$$
Gaussian Process Regression (GPR) was selected as the surrogate model. A GP is a collection of random variables, any finite number of which have a joint Gaussian distribution. It is fully specified by a mean function \(m(\mathbf{x})\) and a covariance (kernel) function \(k(\mathbf{x}, \mathbf{x’})\). For the training data, the outputs \(\mathbf{y}\) are assumed to follow:
$$
\mathbf{y} \sim \mathcal{N}(\mathbf{m}, \mathbf{K})
$$
where \(\mathbf{K}\) is the covariance matrix with entries \(K_{ij} = k(\mathbf{x_i}, \mathbf{x_j}) + \sigma_n^2\delta_{ij}\), and \(\sigma_n^2\) represents the noise variance. A common kernel is the Radial Basis Function (RBF):
$$
k(\mathbf{x_i}, \mathbf{x_j}) = \sigma_f^2 \exp\left(-\frac{1}{2l^2} ||\mathbf{x_i} – \mathbf{x_j}||^2\right)
$$
Here, \(\sigma_f^2\) is the signal variance and \(l\) is the length-scale parameter. The power of GPR lies in its Bayesian predictive capability. For a new test input \(\mathbf{x_*}\), the predictive distribution for the output \(y_*\) is also Gaussian:
$$
p(y_* | \mathbf{x_*}, \mathcal{D}) = \mathcal{N}(\mu_*, \sigma_*^2)
$$
with predictive mean and variance given by:
$$
\mu_* = \mathbf{k}_*^T \mathbf{K}^{-1} \mathbf{y}, \quad \sigma_*^2 = k(\mathbf{x_*}, \mathbf{x_*}) – \mathbf{k}_*^T \mathbf{K}^{-1} \mathbf{k}_*
$$
where \(\mathbf{k}_*\) is the vector of covariances between the test point and all training points. This provides not just a point prediction (\(\mu_*\)) but also a measure of uncertainty (\(\sigma_*\)), which is invaluable for guiding optimization. Before training, the input and output data were normalized using z-score standardization to ensure numerical stability.
3. Optimization via Genetic Algorithm
With the trained GPR model acting as a fast, analytical proxy for the simulation (evaluating in milliseconds vs. hours), an optimization algorithm can efficiently search the 4-dimensional parameter space. A Genetic Algorithm (GA) was employed for this purpose. The GA is a metaheuristic inspired by natural selection. It operates on a population of candidate solutions (chromosomes), each encoding a set of process parameters \([T_p, T_m, t_f, P_h]\). The fitness of each candidate is evaluated as the predicted porosity volume from the GPR model (with a negative sign for minimization). The algorithm iteratively applies selection, crossover (recombination), and mutation operators to evolve the population toward regions of higher fitness (lower porosity). The workflow is as follows:
1. Initialization: Generate a random population of N candidate parameter sets within the defined bounds.
2. Fitness Evaluation: Use the GPR surrogate model to predict porosity for each candidate.
3. Selection: Select parent candidates for reproduction, favoring those with better fitness.
4. Crossover & Mutation: Create offspring by combining parameters from parents and introducing random changes.
5. Replacement: Form a new generation from the best parents and offspring.
6. Termination: Repeat steps 2-5 until a stopping criterion (e.g., number of generations) is met.
Results and Discussion
1. Insights from Numerical Simulation
The initial numerical simulations provided critical physical insights into the aerospace casting process. The filling pattern was smooth and sequential, with the liquid metal front rising steadily in the cavity. Solidification analysis confirmed a largely directional pattern, progressing from the thin upper sections and outer walls towards the thicker feeder gates at the bottom, which is essential for effective pressure-fed feeding. The predicted shrinkage porosity consistently localized in the upper sections of the vertical gates and at isolated hot spots within the cylinder wall junctions, validating the defect prediction capability of the mechanistic model. The results from the 9 orthogonal design simulations served as the foundational dataset, as shown below.
| Run # | Tp (°C) | Tm (°C) | tf (s) | Ph (kPa) | Porosity Volume (cm³) |
|---|---|---|---|---|---|
| 1 | 700 | 250 | 8 | 180 | 2.975 |
| 2 | 700 | 275 | 10 | 200 | 2.301 |
| 3 | 700 | 300 | 12 | 220 | 1.956 |
| 4 | 740 | 250 | 12 | 200 | 2.079 |
| 5 | 740 | 275 | 8 | 220 | 2.132 |
| 6 | 740 | 300 | 10 | 180 | 2.168 |
| 7 | 780 | 250 | 10 | 220 | 2.629 |
| 8 | 780 | 275 | 12 | 180 | 1.962 |
| 9 | 780 | 300 | 8 | 200 | 2.596 |
A preliminary range analysis on this orthogonal data indicated that, within the studied bounds, holding pressure (Ph) had the most significant influence on reducing porosity, followed by pouring temperature (Tp), mold temperature (Tm), and filling time (tf). The optimal combination from this limited set was Ph=220 kPa, Tp=740°C, Tm=275°C, tf=12 s.
2. Performance of the Machine Learning Surrogate
The GPR model was trained on the 9 data points from Table 3. Its predictive accuracy was then rigorously tested on the remaining 72 parameter combinations from the full-factorial design, which were simulated independently to serve as a test set. The performance was exceptional, demonstrating the GPR’s ability to learn the complex, non-linear relationship between process parameters and defect volume from very limited data. A key advantage of GPR is the quantification of prediction uncertainty. For the test set, the mean standard deviation (\(\sigma_*\)) of the predictions was approximately 0.0015, and the 95% confidence interval was within ±0.0029 of the predicted mean. The relative error between the GPR-predicted porosity and the actual simulation result averaged a remarkably low 1.9% across all test cases. This high accuracy and calibrated uncertainty make the surrogate model a reliable tool for optimization.
| Tp (°C) | Tm (°C) | tf (s) | Ph (kPa) | Simulated Porosity (cm³) | GPR Predicted Porosity (cm³) |
|---|---|---|---|---|---|
| 720 | 260 | 9 | 190 | 2.451 | 2.438 |
| 760 | 290 | 11 | 210 | 2.187 | 2.201 |
| 710 | 280 | 10 | 205 | 2.112 | 2.098 |
The most significant benefit is computational speed. A single high-fidelity simulation required approximately 3 hours of computational time. In stark contrast, evaluating the trained GPR model for a new set of parameters took less than 1 millisecond. This speedup of over 10 million times is what enables the exhaustive and iterative search performed by the Genetic Algorithm.
3. Optimal Process Design via Surrogate-Guided GA
The Genetic Algorithm was deployed to minimize the predicted porosity volume, using the GPR model as the fitness evaluator. The GA was not constrained to the discrete levels used in the orthogonal or full-factorial designs but could explore any continuous value within the parameter bounds. After evolution over multiple generations, the GA converged to an optimal solution that surpassed the best design found by the traditional orthogonal analysis. The algorithm-proposed optimum was:
- Pouring Temperature (Tp): 762.3 °C
- Mold Temperature (Tm): 288.0 °C
- Filling Time (tf): 10.1 s
- Holding Pressure (Ph): 196.6 kPa
A final, confirmatory high-fidelity simulation was run using this ML-optimized parameter set. The result validated the prediction: the porosity volume was reduced to 1.76 cm³, which was significantly lower than the 1.96 cm³ from the best orthogonal design combination. This demonstrates that the ML-GA framework can effectively navigate the parameter space to find superior solutions that might be missed by conventional design-of-experiment methods restricted to discrete levels. The optimal parameters balance a relatively high pouring temperature (for fluidity) with a moderately high mold temperature and a specific filling time to establish a favorable thermal gradient, all under a sufficient but not necessarily maximum holding pressure.
Conclusion
This study successfully demonstrates a robust, data-driven framework for the design and optimization of low-pressure die casting processes for complex aerospace casting components. The integration of physics-based numerical simulation, Gaussian Process Regression, and Genetic Algorithms creates a powerful synergy. The mechanistic simulation provides accurate, trustworthy data to train on, capturing the essential physics of mold filling, solidification, and defect formation. The Gaussian Process model efficiently learns the underlying input-output relationships from limited data, providing instant, uncertainty-aware predictions. Finally, the Genetic Algorithm leverages this fast surrogate to perform a global search for the optimal process parameters, free from the grid constraints of traditional methods.
The key findings are:
1. For the cylindrical thin-walled aerospace casting studied, holding pressure is the most influential parameter in reducing shrinkage porosity, followed by pouring temperature.
2. A Gaussian Process Regression model can predict porosity volume with high accuracy (~1.9% mean error) even when trained on a very small dataset (9 points), offering a speed advantage of over 10^7 compared to full simulation.
3. The surrogate-guided Genetic Algorithm identified a process parameter set (Tp=762.3°C, Tm=288.0°C, tf=10.1s, Ph=196.6 kPa) that reduced predicted porosity by approximately 10% compared to the best solution found via classical orthogonal design.
4. This ML-based approach represents a paradigm shift towards rapid, intelligent process design for aerospace casting, significantly reducing the reliance on costly trial-and-error or exhaustive simulation campaigns. It forms a critical component for the development of digital twins and smart foundry systems, enabling real-time process adjustment and closed-loop quality control in advanced manufacturing environments.