Abstract
In this chapter, many of the concepts introduced in the previous chapters are applied to the uncertainty quantification of the lifetime prediction of turbine blades, generated by the uncertainty of the temperature loading field. A complete reduced-order model workflow is detailed, and the numerical experiments make use of the codes Mordicus and genericROM introduced in Chap.
4.
Introduction
Predicting the fatigue life of high-pressure (HP) turbine blades in aircraft engines is a significant computational challenge. These components operate under extreme thermomechanical loads, including centrifugal forces, high-pressure gas, and temperatures that can exceed the material's melting point. To accurately predict crack initiation, engineers must simulate viscoplastic behavior over complex loading cycles.
Standard industrial tools like Abaqus require high-fidelity simulations that can take several weeks for a single run. This latency makes Uncertainty Quantification (UQ)—which necessitates thousands of Monte Carlo scenarios to account for thermal loading variations—practically impossible. This paper presents a "ROM-net" surrogate model designed to be fast enough for many-query analysis while maintaining industrial safety standards.
Methods: The ROM-net Architecture
The authors employ a ROM-net architecture, which functions as a dictionary of specialized local models rather than a single monolithic approximation. The workflow consists of several sophisticated stages:
- Clustering (k-medoids): Training data is partitioned into clusters based on mechanical response, with each cluster assigned its own local Reduced Order Model (ROM).
- Model Order Reduction (POD-ECM): Proper Orthogonal Decomposition (POD) is used for data compression, while the Empirical Criterion Method (ECM) provides hyper-reduction (operator compression).
- Feature Selection (mRMR): A geostatistical "minimum Redundancy Maximum Relevance" algorithm identifies 11 critical nodes out of millions to serve as inputs for the classifier.
- Classification (Logistic Regression): A classifier recommends the optimal local ROM based on temperature data at the selected nodes.
- Gappy Surrogates (Multi-task Lasso): Full-field stress and strain data are reconstructed from limited integration points using a Lasso-based metamodel, reducing bias compared to standard Gappy-POD.
Datasets and Numerical Experiments
The study utilizes a dataset of 200 high-fidelity Finite Element (FE) solutions generated using Z-set software. The mesh consists of approximately one million nodes, providing the necessary resolution for industrial applications.
- MaxProj LHS DoE (80 points): A "Maximum Projection" Latin Hypercube Design used to build the dictionary of local ROMs.
- Sobol’ DoE (120 points): A sequence used to train the classifier and the Gappy surrogates.
- Verification Set: 20 additional Monte Carlo draws used to validate the final ROM-net's accuracy.
Key Results and Insights
The ROM-net approach achieved a 636x speedup during the exploitation phase compared to a highly optimized domain-decomposition FEM solver (AMPFETI). This allowed for 1,008 Monte Carlo simulations to be completed in just 2 hours and 48 minutes.
Accuracy remained high, with relative errors for dual quantities of interest (like Von Mises stress and accumulated plastic strain) between 1% and 2%. Crucially, the model perfectly predicted the locations of maximum stress and strain, which are vital for identifying failure points.
Beyond speed, the method is "green," reducing energy consumption by a factor of 636. A key physical insight was the importance of geodesic distance over Euclidean distance for feature selection on thin structures like blades, ensuring that uncorrelated temperatures on opposite sides of the blade (intrados vs. extrados) were handled correctly.
Limitations and Future Work
The current study is limited to a single loading cycle, predicting the onset of plastic shakedown but not yet performing long-term cyclic extrapolation. Additionally, the assumption of linear thermal behavior for fluctuations may need refinement for highly nonlinear regimes. Future iterations could also incorporate geometric uncertainties and variations in material properties, such as crystal orientation in monocrystalline blades.
Summary updated