CAE Performance Prediction Using Machine Learning Model Based On Historical Data

Modal Frequencies Prediction for Oil Pan using ShapeAI and Deep Learning Model

Machine Learning applications are developed to disrupt product design methodology across all industries. Every design engineer would like to optimize his design at the concept stage only considering a few critical and essential load cases. The major challenge for the design engineer has not much simulation expertise required to prepare the CAE model, apply material properties, load case, solve and post-process to understand the CAE performance. Even, when the engineer has CAE expertise, it will take a considerable amount of time to prepare the CAE model, solve and post-process it.

Machine learning models trained based on historical simulation data can be leveraged to predict the simulation results (stress/strain/ displacement/modal frequencies/fatigue life etc.) for a new design, without going through traditional CAE simulation workflow. This will help the design engineer to understand the effect of each geometry/thickness/material parameter change influencing the CAE performance so that he can generate use those insights in the first stage of design, this will reduce the number of design iterations. A data scientist has to extract design parameters from the 3D design model for building machine learning models, but it is very difficult to extract the design parameters when designs are non-parametric and topologically different from each other. We have come with a methodology to handle those situations, to capture the geometric shape of any part/assembly with a unique shape vector using shapeAI to represent that geometry for machine learning model building.

We have developed some automation scripts on HyperWorks to efficiently extract the shape vector from the geometry/mesh model, extract the thickness and material property from the finite element solver deck, extract CAE responses from the solver result file and finally, it will create tabular data loaded for the training machine learning model with our KnowledgeStudio. We have tried different supervised learning models (Regressing, Classification, Deep Learning, Ensemble Tree) based on historical simulation data (with variations in geometry, material parameters, property parameters) for a component/subsystem level different load cases (structural, modal, fatigue etc.). The machine learning scored model will be used to predict the CAE response for any new design model. This same methodology can be extended to multiple domains and load cases to predict different results, which will come in future blogs. This will significantly reduce the design iterations and improves the product development process to achieve the final design.

This is the standard approach followed:

We created 3000 topologically different oil pan designs as shown below for this study. We used Altair Hyperworks to generate the finite element model for Modal analysis and solved the model in Altair Optistruct solver.

We used HyperWorks and KnowledgeStudio platform for this study as mentioned below.

We trained three machine learning models for this study:

• Linear regression model (learning rate as 0.001 and gradient descent as our optimizer)
• Random forest regression (Number of Trees 500, Max Tree Depth 5, Parametric Measure Entropy Variance)
• Deep learning model (Neural Network Model with Number of Hidden Layers 3, Number of Neurons 65/120/65, Optimizer Conjugate Gradient, Number of Iterations 2000)

 As per the below table different machine model behavior, Deep Learning model convincingly performed better in both the test and train sets, with very good MSE value.

 Please refer to this video for details on data generation and the machine learning model building process.