A hybrid Deep Learning-assisted Finite Element technique to predict dynamic failure evolution in advanced ceramics (DeLFE)

点击翻译按钮获取中文摘要

Engineers use computer tools to design complex structural components, be they airplane wings or high-rise buildings. The most common technique underneath such computer tools is finite element analysis (FEA). For not-so-demanding structural applications, simple linear elastic finite element models are sufficient to conduct the design process. However, when it comes to advanced structural applications such as gas turbine blades or nuclear reactor components, they experience severe loading conditions, which are high temperatures and/or high-pressure dynamic loads (impact). For such high-temperature applications, advanced structural ceramics and their composites are the natural choices of materials. Designing these structures made of ceramics against impact is non-trivial as it involves complex damage evolution mechanisms. Hence, engineers conduct a variety of experiments and large-scale computational (FEA) simulations to understand their behaviour. However, these experiments and simulations are time-consuming and costly. Most often, they have to conduct simple linear elastic FEA models which are far less time-consuming but do not account for complex damage mechanisms. Such a limitation naturally results in a sub-optimal design of these advanced structural components. One of the ways to overcome this challenge is to exploit the potential of artificial intelligence (AI) techniques to see if finite element computer simulations can be accelerated and yet reliably predict complex damage mechanisms. The proposed research addresses this industry-relevant problem and aims to develop an AI-driven accelerated FEA tool to simulate dynamic brittle fracture evolution in advanced structural ceramics. When we use AI models for such physical problems, the inherent problem is their reliability as they are often termed as 'black box modes' leading to uncertainties in their predictions. In this perspective, the proposed project takes a hybrid approach whereby data-driven deep learning models are adaptively combined with physics-based traditional FEA models in an integrated framework, thoroughly validated using experimental testing. Such a hybrid strategy aids in achieving accelerated, yet reliable simulations for complex physical problems such as the dynamic damage evolution in structures made of brittle ceramic materials. More specifically, engineers or end-users can specify the level of the desired accuracy depending on the design stage of the structural component. The adaptive simulation framework will then appropriately hybridise deep learning-based predictor and traditional FEA, resulting in an optimal damage prediction balancing the computational costs and accuracy. The global objective is to equip engineers with the necessary simulation tools whereby both accuracy and speed are ensured. This helps in transitioning the current industrial approach of using simplistic models into adopting AI-driven models that captures complex physical mechanisms, ultimately leading to the efficient and safe design of advanced structural components.