Teaching AI to Predict 3D Stress Fields in Gas Turbine Engine Blisk
As a stress engineer, I spend a significant portion of my time working with finite element (FEA) simulations. While running these models, there was a fundamental question that always fascinated me and kept spinning in my mind:
Can we take geometry and operational loads a step further and teach a machine learning model to predict the entire 3D stress distribution across a rotor blisk, one of the most critical and complex components in a gas turbine engine?
This curiosity led me to develop a deep learning-based surrogate model capable of completely bypassing the massive computational burden of traditional solvers. This study is one of the major milestone in bringing artificial intelligence into structural analysis workflows.
Why Stress Matters in Rotor Blisks
A rotor blisk (bladed disk) integrates the blades and the disk into a single component, offering significant aerodynamic and weight advantages in modern gas turbines. However, this integration means the entire structure must endure incredibly harsh operational environments such as comprising massive centrifugal forces, complex aerodynamic gas pressures, and severe thermal gradients. To guarantee structural integrity and prevent fatigue failures, calculating these multi-axial stress fields accurately is absolutely non-negotiable.
To guarantee structural integrity and safety, calculating these stresses accurately is non-negotiable. However, running high-fidelity commercial FEA software for dozens of different operational scenarios and geometric variations is a massive time and hardware hog. When you want to quickly evaluate hundreds of design combinations during the preliminary phases, simulation times quickly become a frustrating bottleneck.
Creating the Dataset and Training the Neural Network
To teach a machine learning model how to navigate these complex, non-linear structural behaviors, I first needed to feed it the right data.
Simulation Data: Using ANSYS Mechanical, I modeled a single sector of a 70-blade rotor blisk by utilizing cyclic symmetry boundary conditions.
Load Scenarios: I designed 10 distinct static loading conditions combining various RPMs, gas pressures, and temperatures to extract my baseline FEA data.
DNN Architecture: With this dataset, I built a Deep Neural Network (DNN) featuring 4 hidden layers with 256 neurons each. The model takes operating parameters (RPM, Pressure, Temperature) as inputs and predicts the radial, tangential, and axial stress components for every single node in the mesh.
The objective was straightforward: instead of solving massive systems of differential equations at runtime, the network would learn the direct, invisible bridge between the operational boundary conditions and the resulting stress fields straight from the data.
The Results: A 13,200x Speedup
Once the model was trained, I put it to the test using entirely unseen loading scenarios that it had never encountered during training. The results were incredibly exciting:
High Accuracy: Even in complex geometric transition zones and fillets where stress gradients are highest, the AI successfully captured the full 3D stress maps. It achieved a maximum localized deviation of 6% to 15% compared to the baseline FEA—which is highly acceptable for early-stage design filtering.
Incredible Speed: While the traditional FEA solver took 66 seconds to compute the nodal stresses for a single load case, the trained DNN completed the exact same prediction in just 5 milliseconds. We are talking about a massive 13,200x speedup.
Storage Revolution: Traditional FEA result files took up about 34.25 MB on the disk. In stark contrast, the trained DNN model file was only 363 KB. This makes the model incredibly lightweight and easy to deploy in real-time environments.
What Does This Mean for Engineering Workflows?
Just like I’ve argued before, I do not believe artificial intelligence will replace finite element analysis anytime soon. In safety-critical sectors like aerospace and power generation, high-fidelity, physics-based simulations will always remain the gold standard for final verification and certification.
However, when we position AI as a powerful surrogate model, the entire engineering workflow changes. This approach opens the door to:
Running real-time Digital Twins to monitor components on the fly.
Screening thousands of design optimization configurations in seconds without hitting a simulation bottleneck.
Implementing Structural Health Monitoring (SHM) systems that use live sensor data to track component life live.
At the end of the day, these AI tools don't replace engineers. Instead, they free us from spending hours waiting for simulations to finish, allowing us to spend more time doing what we do best: solving actual engineering problems. Rather than viewing FEA and AI as competing technologies, we should see them as complementary partners working together to make engineering faster, smarter, and far more efficient.
Reference: Kortag, U. (2025). Artificial Intelligence-Based Stress Prediction for Rotor Blisk in Gas Turbine Engines. Journal of Aviation Research, 7(2), 149–176.
Yorumlar
Yorum Gönder