Evaluating material models in Finite Element (FE) simulations is computationally expensive. Recently, Machine Learning (ML) techniques have been explored for accelerating elastoplastic algorithms. One such method includes replacing a part of the algorithm with an ML model which is called the “hybrid” approach. One of the most commonly used algorithms for ductile materials is the J2-based von Mises hardening elastoplasticity. To improve the performance of this model, an ML-based hybrid algorithm was sought. In this algorithm, the expensive iterative plastic correction step was replaced with a single-step prediction from a SINDY-inspired sparse nonlinear regression model.
Structural analysis of mechanical components, such as predicting the deformation behavior of sheet metal or assessing the crash safety of a vehicle, typically relies on finite element analysis (FEA). One critical aspect influencing the quality of these simulations are the material models that describe the relationship between strains and stresses. However, the development and selection of the most appropriate models is a significant challenge that involves costly and time-consuming testing and calibration procedures.
The industrial sector uses artificial intelligence (AI) in many ways. E.g. anomaly detection to identify and examine abnormal behavior of machines, such as voltage and current fluctuation. To develop self driving cars AI is used to perform segmentation of the environment to navigate the vehicle and make decisions, preferably in real-time. Quantum computers are already being used for special machine learning processes, achieving, in some instances, better results than a regular machine learning algorithm. This paper will elaborate on the upsides of a machine learning model consisting of a hybrid between a quantum machine learning (QML) algorithm and a classical machine learning algorithm.
Machine Learning (ML) driven material models have been investigated for some time now. The respective ML models can be trained outside of the Finite Element solvers by means of given strain paths and corresponding stress results which have either been recorded from simulations, drawn from distributions, or even measured from hardware tests. The trained models can be easily evaluated regarding their performance based on an-other set of strain paths with results that have not been presented to the model during training. When a model has reached a promising prediction accuracy on the validation data, the natural next step is to test it in a finite element simulation by integrating the trained model as a user material.
A method is presented for generating training data from FEM meshes based on element prop-erties. Learning at the element level is often indicated when larger structures, especially parts, are too inhomogeneous in size or geometry. At the element level, quality metrics and other infor-mation are gathered from the neighborhoods of elements, just as convolutional neural networks for image processing gather information from the neighborhoods of pixels. But in general, neigh-borhoods of elements are not as well structured as pixels in images. Instead, they form irregular graphs which cannot be processed by standard Neural Network (NN) architectures directly.
Accelerating elastoplastic material models with spare nonlinear regression: A hybrid approach