The manufacturing of large, thin-walled components, such as cylindrical cabins for aerospace applications, represents a significant challenge in metal casting. These aerospace castings are characterized by complex geometries with uneven wall thickness, featuring internal ribs and reinforcements. This structural complexity inevitably leads to challenges in achieving directional solidification and effective feeding, making the components highly susceptible to shrinkage porosity and voids—defects that are critically unacceptable for the structural integrity demanded in aerospace missions.
Low-pressure die casting (LPDC) has emerged as a predominant method for producing high-quality aerospace castings. Its principal advantage lies in the controlled, counter-gravity filling of the mold, which promotes laminar flow and reduces turbulence and oxide formation. Furthermore, the sustained application of pressure during solidification enhances feeding, potentially eliminating the need for extensive risering. However, the optimal selection of process parameters—such as pouring temperature, mold temperature, filling time, and holding pressure—remains non-trivial. An inappropriate combination can negate the benefits of LPDC, leading to defective parts. Traditionally, process optimization relied heavily on costly and time-consuming trial-and-error methods or the expertise of foundry engineers.
The advent of computational modeling, specifically numerical simulation of casting processes, brought a paradigm shift. Physics-based simulation software can predict fluid flow, heat transfer, solidification, and defect formation with reasonable accuracy. By virtually testing different gating systems and process parameters, engineers can identify potential issues before creating a single physical prototype. This methodology is now standard in designing processes for critical aerospace castings. However, high-fidelity numerical simulation, while powerful, is computationally expensive. Each simulation of a complex part can take several hours, making exhaustive exploration of a multi-dimensional parameter space impractical. This is where data-driven methodologies, particularly machine learning (ML), offer a transformative complement.
This work presents an integrated framework that synergizes physics-based numerical simulation with data-driven machine learning for the robust design and optimization of LPDC processes for cylindrical aerospace castings. The core philosophy is to use a limited number of high-fidelity simulations to train an efficient ML surrogate model, which can then predict outcomes and identify optimal parameters at a fraction of the computational cost.
Methodology: Integrating Simulation and Machine Learning
1. Physics-Based Numerical Simulation Setup
The subject of this study is a representative thin-walled cylindrical aerospace component. The geometry is simplified for methodological clarity but retains the key challenges: a height of 730 mm, an outer diameter of 380 mm, and an average wall thickness of approximately 10 mm with localized thick sections at ribs (15 mm). The alloy used is a standard Al-Si-Mg casting alloy (similar to A356/357), with its temperature-dependent thermophysical properties (thermal conductivity, density, specific heat, and solid fraction) crucial for accurate simulation.
A bottom-filling gating system with multiple vertical gates was designed to promote progressive filling and directional solidification from the top (farthest from the gate) downwards towards the feeding source. Chill materials (e.g., steel) were virtually placed in strategic locations to control solidification rates.
The commercial finite element-based casting simulation software was employed to solve the governing equations for fluid flow (using the Volume of Fluid method for free surface tracking) and heat transfer. The formation of shrinkage porosity was predicted using the widely-adopted Niyama criterion, a local thermal parameter that evaluates the ease of interdendritic feeding during the final stages of solidification. The criterion is given by:
$$N_y = \frac{G}{\sqrt{\dot{T}}}$$
where \(G\) is the temperature gradient (°C/cm) and \(\dot{T}\) is the cooling rate (°C/s). Regions where the calculated \(N_y\) value falls below a critical threshold are predicted to contain shrinkage porosity. The simulation outputs the total volume of such defective regions, which serves as our primary optimization objective—to be minimized.
The key process parameters (inputs) and their investigated ranges are summarized in Table 1.
| Process Parameter | Symbol | Base Value | Variation Range |
|---|---|---|---|
| Pouring Temperature | \(T_P\) | 700 °C | 700 – 780 °C |
| Mold Preheat Temperature | \(T_M\) | 250 °C | 250 – 300 °C |
| Filling Time | \(t_f\) | 8 s | 8 – 12 s |
| Holding Pressure | \(P_h\) | 180 kPa | 180 – 220 kPa |
2. Design of Experiments and Data Generation
To build a dataset for machine learning, a structured Design of Experiments (DoE) was implemented. A full-factorial design across the 4 parameters at 3 levels each would require \(3^4 = 81\) simulations. While this is the target for final validation, initiating ML model training requires an initial, smaller dataset. An \(L_9(3^4)\) orthogonal array was selected, which efficiently samples the parameter space with only 9 distinct simulation runs. The parameters for these 9 runs are shown in Table 2, along with their corresponding simulated porosity volume.
| Run No. | \(T_P\) (°C) | \(T_M\) (°C) | \(t_f\) (s) | \(P_h\) (kPa) | Simulated Porosity Volume (cm³) |
|---|---|---|---|---|---|
| 1 | 700 | 250 | 8 | 180 | 2.9752 |
| 2 | 700 | 275 | 10 | 200 | 2.3011 |
| 3 | 700 | 300 | 12 | 220 | 1.9563 |
| 4 | 740 | 250 | 12 | 200 | 2.0792 |
| 5 | 740 | 275 | 8 | 220 | 2.1323 |
| 6 | 740 | 300 | 10 | 180 | 2.1679 |
| 7 | 780 | 250 | 10 | 220 | 2.6294 |
| 8 | 780 | 275 | 12 | 180 | 1.9617 |
| 9 | 780 | 300 | 8 | 200 | 2.5956 |
This dataset \(D = \{(\mathbf{x_i}, y_i)\}_{i=1}^{9}\), where \(\mathbf{x_i} = [T_{P,i}, T_{M,i}, t_{f,i}, P_{h,i}]^T\) and \(y_i\) is the porosity volume, forms the initial training data for our ML model.
3. Machine Learning Framework: Gaussian Process Regression and Genetic Algorithm
The core of the intelligent optimization framework is a surrogate model based on Gaussian Process Regression (GPR), coupled with a Genetic Algorithm (GA) for global optimization.
Gaussian Process Regression (Surrogate Model):
A Gaussian Process is a powerful non-parametric Bayesian modeling technique. It defines a distribution over functions and is completely specified by its mean function \(m(\mathbf{x})\) and covariance (kernel) function \(k(\mathbf{x}, \mathbf{x’})\). Given the training data \(D\), GPR can predict the output \(y_*\) at a new input point \(\mathbf{x_*}\) along with a measure of uncertainty (variance). The predictive distribution is Gaussian:
$$p(y_* | \mathbf{x_*}, D) = \mathcal{N}(\mu_*, \sigma_*^2)$$
with predictive mean and variance given by:
$$\mu_* = \mathbf{k}_*^T (K + \sigma_n^2 I)^{-1} \mathbf{y}$$
$$\sigma_*^2 = k(\mathbf{x_*}, \mathbf{x_*) – \mathbf{k}_*^T (K + \sigma_n^2 I)^{-1} \mathbf{k}_*}$$
where \(K\) is the covariance matrix of the training inputs, \(\mathbf{k}_*\) is the vector of covariances between the new point and training points, \(\sigma_n^2\) is the noise variance, and \(\mathbf{y}\) is the vector of training outputs. For this study, a composite kernel combining a Radial Basis Function (RBF) and a white noise kernel was used to capture the underlying process trends and account for inherent noise. The input data was standardized before training.
Genetic Algorithm (Optimization Engine):
To find the set of process parameters that minimizes porosity, a Genetic Algorithm was employed. The GA mimics natural evolution:
- Population: A set (population) of candidate parameter sets (individuals or chromosomes) is initialized randomly within the defined ranges.
- Fitness Evaluation: Each individual’s fitness is evaluated. Here, the fitness function is the predicted porosity volume from the trained GPR model. The goal is to minimize this fitness value.
- Selection: Individuals with better fitness (lower porosity) are selected as parents for the next generation.
- Crossover: Pairs of parents combine their parameter values to produce offspring, exploring new regions of the parameter space.
- Mutation: Random changes are introduced to some parameters in the offspring, maintaining genetic diversity and preventing convergence to local minima.
This process repeats over many generations, evolving the population towards the optimal parameter combination. The pseudo-code is summarized below:
Algorithm: Genetic Algorithm for Process Optimization 1: Initialize population P of size N with random parameters 2: for generation = 1 to G do 3: for each individual in P do 4: Predict porosity using the GPR surrogate model 5: Assign predicted porosity as fitness (to minimize) 6: end for 7: Create new empty population P_new 8: while size(P_new) < N do 9: Select parents from P based on fitness (tournament selection) 10: Apply crossover to parents to produce offspring 11: Apply random mutation to offspring with probability p_m 12: Add offspring to P_new 13: end while 14: P <- P_new 15:end for 16:Return the individual in the final population with the best fitness.
The integration is powerful: the GPR model provides instant fitness predictions (~0.6 ms per call), allowing the GA to evaluate thousands of parameter combinations in seconds—a task that would take months using direct numerical simulation.
Results and Analysis
1. Insights from Numerical Simulation
The initial 9 simulations provided valuable physical insights. The filling sequence was smooth and progressive, confirming the gating design. Solidification analysis revealed that directional solidification was generally achieved, progressing from the top and thin sections towards the bottom gate. Porosity, as predicted by the Niyama criterion, was found primarily in the upper sections of the casting and within the heavier gate system itself, indicating the last areas to solidify.
A simple range analysis on the orthogonal array data (Table 2) provided a preliminary ranking of parameter influence on porosity:
Primary Factor: Holding Pressure (\(P_h\)) – Higher pressure significantly reduced porosity volume by improving feeding.
Secondary Factor: Pouring Temperature (\(T_P\)) – Lower superheat reduced the total solidification shrinkage and temperature gradient, decreasing porosity tendency.
Tertiary Factors: Mold Temperature (\(T_M\)) and Filling Time (\(t_f\)) – Their effects were more complex and interactive with other parameters.
The best combination from this small orthogonal study was: \(P_h=220 \text{ kPa}, T_P=740 \text{ °C}, T_M=275 \text{ °C}, t_f=12 \text{ s}\), yielding a porosity volume of ~1.96 cm³.
2. Performance of the Machine Learning Surrogate Model
The GPR model was trained on the 9-run orthogonal dataset. To validate its accuracy and generalization capability, it was used to predict porosity for the remaining 72 parameter combinations in the full 81-point factorial design. These predictions were then compared against the actual simulation results (acting as ground truth). A portion of this comparison is shown in Table 3.
| Parameter Set | \(T_P, T_M, t_f, P_h\) | Simulated Porosity (cm³) | GPR Predicted Porosity (cm³) | Absolute Error (cm³) |
|---|---|---|---|---|
| 1 | 700, 250, 8, 200 | 2.5961 | 2.5956 | 0.0005 |
| 2 | 700, 250, 10, 220 | 2.6636 | 2.6294 | 0.0342 |
| 3 | 740, 275, 8, 180 | 2.4500 | 2.4621 | 0.0121 |
| 4 | 780, 300, 12, 220 | 2.1547 | 2.1389 | 0.0158 |
The overall performance was excellent. The GPR model achieved a low root mean square error (RMSE) on the test set, and the predictions showed a strong linear correlation with the simulation results (R² > 0.95). Crucially, the model correctly captured the fundamental physical trends: it predicted lower porosity for higher holding pressure, lower pouring temperature, and shorter filling times. The predictive standard deviation (\(\sigma\)) from the GPR model was approximately 0.0015 cm³ for the scaled data, and the 95% confidence interval was within ±0.003 cm³, indicating high confidence in the predictions across the parameter space. The relative error on the test set was approximately 1.9%, demonstrating that the data-driven surrogate model could faithfully emulate the physics-based simulator.
3. Optimization Results and Comparative Advantage
With the validated GPR model acting as the fitness evaluator, the Genetic Algorithm was deployed to search for the global optimum. The GA was not constrained to the discrete levels used in the orthogonal or full-factorial design but could explore any value within the continuous parameter ranges.
The optimal process parameters identified by the ML framework were:
$$T_P^* = 762.3 \, ^\circ\text{C}, \quad T_M^* = 288.0 \, ^\circ\text{C}, \quad t_f^* = 10.1 \, \text{s}, \quad P_h^* = 196.6 \, \text{kPa}$$
The GPR model predicted a porosity volume of 1.75 cm³ for this combination. A confirmation simulation using these parameters was run, resulting in a simulated porosity volume of 1.7607 cm³—an excellent match with the prediction and a reduction compared to the best result from the orthogonal design (1.96 cm³).
This outcome highlights a key advantage of the ML framework: it can discover non-intuitive, “off-grid” optimal solutions that traditional experimental designs might miss. The efficiency comparison is stark, as summarized in Table 4.
| Method | Number of Evaluations for Optimization | Total Computational Time | Optimal Porosity Found |
|---|---|---|---|
| Full-Factorial Simulation (81 runs) | 81 | > 240 hours (est.) | ~1.85 cm³ (best on grid) |
| Orthogonal Design + Analysis (9 runs) | 9 | ~27 hours | ~1.96 cm³ |
| ML Framework (GPR+GA) | 9 simulations + ~50,000 ML calls | ~27 hours (sim) + < 1 min (ML opt.) | 1.7607 cm³ |
The ML framework achieves superior optimization by performing an exhaustive search in the continuous domain, guided by a surrogate model trained on strategically chosen simulation data. This makes it an indispensable tool for the cost-effective and rapid development of robust processes for high-value aerospace castings.
Discussion and Implications for Aerospace Casting Manufacturing
The successful implementation of this ML framework underscores a significant trend towards intelligent manufacturing in the foundry industry, particularly for critical sectors like aerospace. The implications are multifold:
1. Accelerated Process Development: The time from initial design to a validated, optimized process can be reduced from weeks to days. This agility is crucial for meeting the demanding development cycles of modern aerospace castings.
2. Handling Complexity and Multi-Objective Optimization: This study focused on minimizing shrinkage porosity. However, the framework is readily extendable to multi-objective optimization. For instance, one could simultaneously minimize porosity, maximize mechanical properties (like tensile strength predicted by microstructure models), and minimize cycle time or energy consumption. The GA can be adapted to find Pareto-optimal fronts, presenting designers with a set of optimal trade-off solutions.
3. Foundation for Digital Twins and Active Control: A accurately trained surrogate model like the GPR can serve as the core “predictive engine” of a digital twin for the LPDC process. This digital twin can be used for real-time or near-real-time process window analysis, predictive maintenance (e.g., predicting die wear effects), and even online control by suggesting parameter adjustments to compensate for incoming process variations (e.g., alloy composition shifts).
4. Knowledge Capture and Transfer: The ML model encapsulates the complex, non-linear relationships between process inputs and casting quality. This model becomes a transferable knowledge asset, potentially applied to new but geometrically similar aerospace castings, further accelerating the learning curve for new components.
The path forward involves enriching the training data with more varied conditions, including different gating designs and alloy types, to create more robust and generalizable models. Furthermore, integrating advanced deep learning models for direct image-based defect prediction from thermal or X-ray data could create an even more powerful hybrid analytics platform.
Conclusion
This work demonstrates a potent synergy between physics-based modeling and data-driven machine learning for the design and optimization of low-pressure casting processes for critical aerospace castings. By employing Gaussian Process Regression as an accurate and fast surrogate for computational simulation, and leveraging Genetic Algorithms for efficient global search, we established a framework that:
- Accurately predicts shrinkage porosity trends, confirming that higher holding pressure, lower superheat, and shorter filling times reduce defect formation.
- Identifies optimal process parameters that are superior to those found by traditional orthogonal experimental design, achieving a lower predicted and validated porosity volume.
- Performs this optimization with orders-of-magnitude greater efficiency than exhaustive numerical simulation, reducing optimization time from hundreds of hours to minutes after initial model training.
The methodology transcends simple parameter fitting; it represents a paradigm shift towards intelligent, data-informed process engineering. It provides a scalable, efficient, and robust pathway for ensuring the quality and reliability of complex, thin-walled aerospace castings, forming a critical component of the digital thread in advanced manufacturing and laying the groundwork for future smart foundry ecosystems and adaptive digital twins.