Development of a methodology for plausibility checks for linear structural mechanic finite element simulations using Deep Learning

在当前的工业环境中，伴随线性有限元模拟的设计通常由产品开发人员进行，而不仅仅是由具有多年专业经验的计算工程师进行。这将导致产品开发过程中的频繁迭代，并可能导致基于未充分验证的结果的错误决策。线性结构-力学有限元模拟的自动合理性检验是支持产品开发人员的重要方法。卷积神经网络（CNN）和机器学习方法的使用代表了识别数据相关性和构建具有高预测质量的模型的巨大潜力。在申请人的准备工作中，可以证明使用深度学习和cnn进行有限元计算的合理性检查是可能的。然而，需要通过调整参数来提高人工神经网络的预测质量，并证明该方法在新的仿真中的应用。此外，应研究有限元模拟的局部区域，特别是检测奇异性等数值误差。本课题的目的是在投影法和奇异识别的基础上，建立一种相似线性结构力学有限元模拟的合理性检验方法。此外，还需要优化不同cnn和机器学习方法的网络参数，以实现高预测质量的可信性检查。有限元模拟结果不能直接转换为神经网络或机器学习算法，而必须转换为统一的计算机可处理形式。在研究项目的框架内，采用了前期工作中开发的投影方法，该方法使用球形探测器表面将任意模拟转换为均匀大小的矩阵。生成的矩阵包含将模拟分类为可信或不可信的所有相关信息。一个深度学习CNN或SVM进行分类。除了对整个模拟进行分类外，还应检查局部区域。特别是在有限元模拟中，需要检测到奇异点并反馈给用户。通过自动合理性检查，可以在早期阶段自动检测模拟设置中的错误。因此，它代表了在虚拟产品开发中提高仿真质量的巨大潜力。特别是如果FE模拟是由产品开发人员执行的，他们的模拟知识比经验丰富的计算工程师少。将开发一种方法，允许考虑类似的几何形状和线性有限元模拟的模拟边界条件。

In the current industrial environment, design accompanying linear finite element simulations are often carried out by product developers and not exclusively by calculation engineers with several years of professional experience. This leads to frequent iterations in the product development process and can lead to incorrect decisions based on insufficiently validated results. An automatic plausibility check for linear structural-mechanical FE simulations is an important method to support product developers. The use of Convolutional Neural Networks (CNN) and Machine Learning methods represents an enormous potential to identify correlations in data and to build a model with high prediction quality. In the applicant's preparatory work it could be shown that a plausibility check for FE calculations using Deep Learning and CNNs is possible. However, it is necessary to increase the prediction quality of the artificial neural network by adjusting the parameters and to demonstrate the application of the method to new simulations. Furthermore, local areas of the FE-simulation shall be investigated, especially to detect numerical errors like singularities.The aim of the project is to create a method for plausibility checks of similar linear structural-mechanical FE simulations based on the preliminary work on the projection method and singularity recognition. Furthermore, the network parameters of different CNNs and machine learning methods are to be optimized in order to implement a plausibility check with high prediction quality. FE-simulation results cannot be directly transferred to a neural network or machine learning algorithm, but have to be converted to a uniform computer-processable form. Within the framework of the research project, the projection method developed in preliminary work is applied, which uses spherical detector surfaces to transform arbitrary simulations into matrices with uniform size. The generated matrices contain all relevant information to classify a simulation as plausible or implausible. A Deep Learning CNN or SVM does the classification. In addition to the classification of the entire simulation, local areas should also be examined. Especially singularities in FE-simulations shall be detected and accordingly give feedback to the user.With an automatic plausibility check, errors in the simulation setup can be detected automatically at an early stage. It therefore represents an enormous potential for increasing the simulation quality in virtual product development. Especially if FE simulations are performed by product developers who have less simulation knowledge than experienced calculation engineers. A method will be developed which allows to consider similar geometries and simulation boundary conditions of linear FE-simulations.