Component Fatigue Response Prediction With Recurrent Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Macroscale finite element methods for simulating mechanical responses are computationally expensive due to the need for full-field microstructural simulations, limiting their practical use in predicting fatigue damage and crack initiation in components.
Innovation Solution
A method combining macroscale simulations with a trained recurrent neural network (RNN) to predict fatigue indicator parameters, replacing computationally intensive crystal plasticity finite element methods, allowing for efficient prediction of mechanical responses and crack initiation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If full-field microstructural simulations (crystal plasticity finite element method) are used to determine material behavior at each integration point, then prediction accuracy of fatigue damage and crack initiation is improved, but computational cost and time increase significantly
Solution Approach 1:
The patent applies preliminary action by training a recurrent neural network model in advance using microstructural simulation data. The RNN is pre-trained to learn the relationship between macroscopic deformation and microscopic stress responses, enabling it to predict material behavior without performing computationally expensive microstructural simulations during the actual fatigue analysis. This preliminary training phase transfers the computational burden from the analysis stage to the model preparation stage.
Solution Approach 2:
The patent replaces the crystal plasticity finite element method (a detailed mechanical microstructural simulation system) with a recurrent neural network model. The RNN learns to approximate the complex microstructural behavior through data-driven patterns, substituting the physics-based mechanical simulation with a computational intelligence approach that achieves similar predictive accuracy with significantly reduced computational cost.
2Reliability
If crystal plasticity finite element method is applied at each integration point of macroscale simulation, then microstructural-dependent fatigue response is accurately captured, but device complexity and computational resources increase
Solution Approach 1:
The patent introduces a recurrent neural network model as an intermediary between the macroscale finite element simulation and the microstructural response prediction. The RNN acts as a surrogate model that captures the complex microstructural effects without requiring direct coupling with detailed crystal plasticity simulations at each integration point, thereby simplifying the overall simulation framework while maintaining predictive accuracy.
Solution Approach 2:
The patent creates a computational copy of the microstructural behavior through the recurrent neural network. Instead of performing actual microstructural simulations, the RNN is trained to replicate the stress-strain response and fatigue indicator parameter evolution that would result from crystal plasticity simulations. This copying approach preserves the essential microstructural effects while avoiding the computational complexity of the original simulation method.
3Productivity
If standard finite element method is used for macroscale simulation, then computational efficiency is maintained, but microstructural aspects and early fatigue damage cannot be adequately modeled
Solution Approach 1:
The patent merges the computational efficiency of standard finite element methods with the predictive capability of microstructural simulations by integrating a recurrent neural network into the macroscale simulation framework. The RNN incorporates microstructural knowledge learned from crystal plasticity simulations, enabling the standard FEM to predict fatigue damage and crack initiation with microstructural accuracy while maintaining computational efficiency.
Solution Approach 2:
The patent changes the approach from direct microstructural simulation to using a trained neural network model that predicts microstructural responses based on macroscopic deformation parameters. This parameter-based approach allows the standard finite element method to access microstructural-level predictions by querying the RNN with deformation data, thereby enhancing fatigue prediction capability without sacrificing computational efficiency.
Data Source
Figure 1~2
AI summary
The present invention relates to a computer-implemented method for determining a distribution of a fatigue indicator parameter (FIP(p,t)) of a component after applying a load characteristics, comprising the steps of: - Simulating (S2) the microstructural response using a macroscale simulation applying a finite element method in consecutive time steps, wherein the microstructural response is obtained in a recurrent process of concurrently determining deformation (ε(p,t)) and stress (σ(p,t)) for each integration point until macroscale simulation has converged to a balancing equilibrium; - For each iteration, each time step and each integration point, applying (S3) a trained recurrent neural network (12) based on a deformation (ε(p,t)) increment to obtain the stress (σ(p,t)) and fatigue indicator parameter (FIP(p,t)), wherein the fatigue indication parameter (FIP(p,t)) is derived from an internal state of the recurrent neural network (12).