Digging out Precious Information Hidden Behind Indirect Experimental Data of Fiber-reinforced Composites using Machine Learning Approach

Cheng Qiu, Jinglei Yang*

Hong Kong University of Science and Technology, Hong Kong SAR, China

ABSTRACT: In composites science field, with the guidance of domain knowledge, machine learning has been used as response predictor to generate mechanical and non-mechanical response of composite structure or as inverse design tool to accelerate material or structure design process. Here, we are aiming at using machine learning to dig out the hidden information of composites from indirect data measured by simple experiments. While experimental data is sometimes limited by the insufficient prepared samples, measuring equipment, or lack of ability to show real-time information inside laminates, some information of laminates are hard to be directly shown by experimental data. For example, the in-situ strength of embedded single-ply, load distribution ratio of composite mechanical joint and so on. However, with the help of machine learning model and domain knowledge to enhance understanding of experiment-related properties, it is possible to find out more about this hidden information behind indirect experimental data. The machine-learning based framework (Fig.1) can be either used as an alternative way to obtain properties of composites, or as a design tool for quick selection of material used in structure. Two examples are given. One is that with the incorporation of Finite Fracture Mechanic, our machine learning model can provide as a simple way to acquire crack resistance curve of fiber reinforced laminates by experimental measuring strength of center-cracked laminates. Another one is the layer-wise design of composite structure under a specific stiffness or strength requirement. By combining of neural network systems with finite element model to interpret the indirect inputs, our machine learning model can determine the needed single-ply property and layup design.

Figure 1 Machine learning based framework

Keywords: Machine Learning; Neural Network; Experimental Data; Material Property