Hybrid Deterministic/Statistical Multi-scale Modelling Techniques for 3D Woven Composites

如今，复合材料处于工程革命的前沿，其目标是制造更轻、更可靠、更省油的航空航天结构。先进的复合材料是由长纤维层结合在一起，使用基质形成结构。在航空航天应用中使用的最常见的纤维类型是碳纤维与环氧树脂基体的结合。最近引入了其他类型的纤维/基体，例如：用于高温应用的陶瓷基复合材料和用于耐磨/耐冲击的金属基复合材料。然而，所有类型的复合材料之间的共同点是它们都基于纤维层。根据定义，图层是2D的。因此，所有传统的复合材料在第三个方向上都难以承受直接载荷。虽然2D复合材料为设计师提供了明显的优势，这些优势来自于纤维的优越性能和定制纤维方向或组合不同纤维类型的灵活性，但厚度性能仍然是限制其充分发挥潜力的致命弱点。3D复合材料是解决这些问题的可行方案，因为它们是由三维编织的纤维制成的。这些材料表现出很大的希望，因为它们可以通过厚度承受直接载荷，并可以抵抗冲击事件。然而，使用3D复合材料会带来一系列建模挑战，这使得工程师无法充分利用这些材料。传统上，为了理解一种新材料的行为，工程师和科学家需要测试材料的样品来表征其行为。然后将这种特征行为包含在可以预测由这种材料制成的结构行为的数学模型中。这些结构模型被用作设计工具。这种传统方法不适用于3D复合材料。在制造过程中，编织纤维的三维网络在拐角处和其他结构特征周围变形，以符合结构几何形状。这反过来意味着光纤网络的每个部分都有不同的架构，因此会有自己的特征行为。因此，简单的材料测试不再能描述材料的性能，需要一种替代方法。该项目旨在训练模型检测结构中存在的3D纤维编织网络中的重复模式。这些可重复的图案将使用非常详细的模型来描述，以了解每个图案在不同的加载条件下以及作为多个结构的一部分如何表现。使用这种方法，将使用无监督机器学习建立一个包含数千个可重复模式及其行为的参数化数据库。在结构尺度上，一个完整的结构的行为可以从重复模式的行为组装而成，而不管它的几何形状。这种方法将允许工程师首次同时设计结构和形成它的3D纤维网络。实现这一目标使我们能够建造更轻、飞行消耗更少燃料、更便宜、生产速度更快的航空结构。使用统计模型来描述结构行为的概念已经出现了一段时间。然而，这些方法一直被认为是一种黑盒解决方案，可以回答结构/材料会发生什么，而不是为什么会发生。在这个项目中，使用了一种混合方法，将统计模型与基于物理的确定性模型相结合。混合方法提供了有关机械性能的信息，以及有关特定行为发生的潜在物理原因。这将使工程师和科学家能够更深入地了解3D复合材料的行为，而不是目前仅通过统计或确定性模型。

Today, composite materials are at the forefront of an engineering revolution targeting lighter, more reliable, and more fuel-efficient aerospace structures. Advanced composites are made from layers of long fibres bound together using a matrix to form the structure. The most common fibre type used in aerospace applications are Carbon Fibres combined with an Epoxy matrix. More recently other types of fibres/matrix are being introduced, such as: ceramics matrix composites for high temperature applications and metal matrix composites for abrasion/ impact resistance. However, something common between all types of composites is that they are based on fibre layers. By definition, layers are 2D. As a result, all conventional composite materials struggle with direct loading in the third direction. While 2D composites provide designers with clear advantages coming from the superior properties of the fibres and the flexibility of tailoring fibre directions or combining different fibre types, through thickness performance remains an Achilles heel that have limited their full potential.3D Composites is a viable solution to these issues as they are made from fibres woven in all three dimensions. These materials show a lot of promise as they can carry direct load through thickness and can resist impact events. However, there are a set of modelling challenges that come with using 3D composites, which have prevented engineers from taking full advantage of these materials. Traditionally, to understand a new material behaviour, engineers and scientists test samples of the material to characterise its behaviour. Then this characteristic behaviour is included in the mathematical models that can predict the behaviour of structures made from this material. These structure models are what is used as design tool. This conventional approach does not work for 3D composites. During manufacturing, the 3D network of woven fibres deforms around corners and other structural features to conform to the structure geometry. This in turn means that the fibre network will have a different architecture for each part of the structure and consequently will have its own characteristic behaviour. As a result, simple material testing is no longer descriptive of the material behaviour and an alternative approach is needed.This project aims to train models to detected repeating patterns that exist in a 3D woven network of fibres across a structure. These repeatable patterns will be characterised using highly detailed models to understand how each pattern behaves under different loading conditions and as part of multiple structures. Using this approach, a parameterised database containing thousands of these repeatable patterns and their behaviour will be built using unsupervised machine learning. On the structure scale, the behaviour of a full structure can be assembled from the behaviour of the repeating patterns forming it regardless of its geometry. This approach will allow engineers, for the first time, to design both the structure and the 3D fibre network forming it simultaneously. Achieving this goal allows us to build aerospace structures that are lighter, consume less fuel to fly, cheaper and faster to produce. The concept of using statistical models for describing structural behaviour have been around for some time. However, these approaches have always been proposed as a black box solution that can give an answer regarding what will happen to a structural/material but not why it happened. In this project, a hybrid approach is used, which combines statistical models with physically based deterministic models. The hybrid approach provides information about the mechanical performance, as well as the underlying physical reasons regarding why a given behaviour happens. This will allow engineers and scientist to understand 3D composites behaviour at a much deeper level than is currently possible by the statistical or deterministic models alone.