Development of a databased model for the prediction of effective mechanical and thermal properties of injection-moulded semi-crystalline thermoplastics by means of an artificial neural network (KNN) taking into account the microstructure

半结晶热塑性塑料局部组分性能的预测是注塑成型中的一个重要挑战。为了确定这些特性作为工艺的函数，通常进行注射成型模拟。然而，没有考虑到的是组件的微观结构，上述两个模拟都发生在宏观尺度上。这意味着构件内部的性能被认为是均匀的，尽管实际上微观结构中存在不均匀性。这种不均匀的微观结构导致注射成型半晶部件的局部热性能和机械性能的不均匀性。因此，在塑料加工研究所（IKV）开发了一个预测微观结构的模型。在一次合作中，还开发了一个额外的模型来模拟所得微观结构的均匀化，该模型模拟了有效的局部特性。目前，均质化的执行伴随着高计算成本，这对于工业4.0应用的性能实时预测来说是不现实的。因此，在本项目的背景下，试图在不影响模拟预测质量的前提下，通过模型缩减的方式，大幅减少均匀化的模拟时间。为此目的，主要审查基于数据的方法的适用性。目的是建立一个数据库模型，通过相似度分析来预测注射成型半结晶热塑性塑料的有效力学性能和热性能。该项目分为四个工作包。第一个工作包包括定义微观结构表征的参数，这些参数可以用作相似性分析的比较变量。在第二个工作包中，使用塑料加工研究所（IKV）开发的SphaeroSim软件建立了模拟微观结构的数据库。另一方面，要创建一个合成结构的数据库，它的生成速度比模拟的微观结构快得多。第三个工作包由两个步骤组成。首先，确定确定的比较值对有效性能的影响。这样就可以整理出不合适的比较值或者定义新的比较值。第二步，将训练不同的神经网络来进行相似度分析。网络类型是递归神经网络和卷积神经网络。为了确保模型的有效性，应该在最后的工作包中进行验证。为此，将神经网络的结果与原始模型的结果进行比较，该模型通过微观结构的均匀化来计算有效性能。

The prediction of the local component properties of semi-crystalline thermoplastics is an important challenge in injection moulding. In order to determine these properties as a function of the process, injection-moulding simulations are often carried out. However, what is not taken into account is the microstructure of the component, the two mentioned simulations take place on the macro scale. This means that the properties within the component are considered as homogeneous, although actually there exists inhomogeneity in the microstructure . This inhomogeneous microstructure causes an inhomogeneity in the local thermal and hence mechanical properties of an injection-moulded semi-crystalline component. Therefore, a model for the prediction of the microstructure was developed at the Institute of Plastics Processing (IKV). In a cooperation an additional model for the homogenization of the resulting microstructure was also developed, which simulates the effective local properties. Currently the execution of the homogenization is accompanied by high computational cost, which is not practical in terms of real-time prediction of properties for Industry 4.0 applications. Therefore, in the context of this proposed project, it is attempted to reduce drastically the simulation time of the homogenization by means of model reduction without worsening the prediction quality of the simulation.. For this purpose data-based approaches are examined for applicability primarily.The aim is to build a databased model, which can be used to predict the effective mechanical and thermal properties of injection-moulded semi-crystalline thermoplastics by means of a similarity analysis. The project is divided into four work packages. The first work package consists of defining parameters for the characterisation of microstructure, which can be used as comparative variables for the similarity analysis. In the second work package, a database of simulated microstructure structures is built up using the software SphaeroSim developed at the Institute of Plastics Processing (IKV). On the other hand, a database of synthetic structures is to be created, which can be generated much more quickly than the simulated microstructures. The third work package consists of two steps. First, the influence of the defined comparative values on the effective properties is to be determined. In this way, inappropriate comparative values can be sorted out or new comparative values can be defined. In the second step, different neural networks from will be trained to perform the similarity analysis. The network types are a Recurrent Neural Networks and a Convolutional Neural Network. To ensure that the validity of the model is guaranteed, it should be validated in the last work package. To this end, a comparison of the results of the neural networks with the results of the original model is aimed at, which calculates the effective properties by a homogenization of the microstructure.