High performance composites (HPC) are a combination of two distinct components namely, fibers and resin, which exhibits completely different properties from fiber and resin. The new (composite) material shows high strength and stiffness and low weight and is stronger alternative to traditional manufacturing materials including steel and aluminum. 
The advantages of HPC materials are accompanied by some challenges related to their implementation into products. The main issues are related to complex mechanical analysis of materials and parts; lack of reliable methods for fatigue and failure prediction and time, temperature and moisture dependent effects. All these factors result in very demanding prediction of failure of the HPC materials and products. Additional problems are associated with increased use of thermoplastic matrices (PES, PEEK, PE, etc.) in composites, since they are more sensitive to temperature and humidity changes compared to conventional crosslinked matrix (e.g. epoxy). 
The structural health monitoring (SHM) systems are used, among others, to measure response of structures made from composite materials. Signal from such system is comprised of two signals: one resulting from change in the geometry (cracks and delaminations), the other coming from change in matrix material properties due to temperature and humidity. In order to distinguish between geometry and matrix material related changes in the response signal of a complex and even noisy signal, one needs a robust tool. This tool should be able to solve in a real time the problem interrelating excitation, material properties (time-, temperature- and humidity-dependent) and structure response. Analytical solution interrelating these factors for a complex geometry and custom excitations does not exist. Numerical solution applied to the complex geometry in presence of noise becomes merely computationally impossible in the real time. 
That is why, I am proposing within this project to utilize a Multilayer Perceptron Neural Network (MLP NN) for determination of the polymer matrix material creep response exposed to different temperatures. The project aims at modeling behavior of polymer material under ramp and harmonic stress loading and influence of the temperature.

Artificially generated data will allow wide set of investigations to address various practical questions, such as effect of data acquisition rate, frequency and amplitude of dynamic loading, noise level effect, effect of number of training data, effect of temperature on the performance of neural network method. Substantial experience of the researcher in determination of temperature dependent mechanical properties will allow refinement and validation of the NN model using experimental creep data.

The ultimate goal of the project is to provide real-time monitoring tool that will be able to filter out the effects of temperature on the matrix from complex signal from the system. The first step towards this goal, represents this proposed project. Within this project the applicant will utilize Multilayer Perceptron Neural Network (MLP NN) for determination of the polymer material creep response exposed to different temperatures. Aim of the project is to model the behavior of polymer material under ramp and harmonic stress loading and influence of the temperature.

The project will cover two different types of excitation: linearly increasing loading and harmonically changing loading. During operation the composites structures in aircraft for example are subjected to various excitations including the harmonic ones (motor vibrations) and impact-like (turbulence). Considering the nature of load during the operation of vehicles and structures, they are mostly subjected to the stress loading and therefore material undergoes creep process. Therefore, within this project, stress loading will be used as a mode of loading.

Research of linearly increasing loading and harmonic loading is separated into 2 Work Packages (WP) but the steps (work tasks – WT) within each of them are similar, though involve different variables. In the first two work tasks, data on shear creep compliance will be generated artificially using a known closed-form solution mimicking behavior of one of the matrix materials utilized in the aerospace industry (epoxy, PE, PEEK, PES, etc.) – details are described in Table 2, see work tasks 1 and 2 (WT1-2). Training data will be generated using different temperatures, noise levels, data acquisition rate and excitation parameters – see WT3-4. Validation after training on artificial data (WT5) and analysis of result (WT6) will be followed by measurements in laboratory and validation with the real measured data on the material chosen in WT1 (WT7).

Current status

Both Work Packages were carried out in parallel as originally planned. For the experimental part of the project that is planned in WT7 the Polyetheretherkethon (PEEK) material has been chosen, based on availability of material and samples of required shapes. The characterization of the material has started earlier that has been originally planned in order to provide the information on material properties earlier. Meanwhile all modelling and neural network trials are carried out on the existing data obtained from previous research works of the CEM research group on Polyphenylene sulfide (PPS), which also falls into group of high-performance polymers.

WP1, related to the material behavior under linearly increasing stress rate is carried out up to the WT4 and WT5 is now in progress. It has been planned to present the first results demonstrating how amount of training data at different temperatures affects the neural network performance at the international conference of Time Dependent Materials, but unfortunately, due to the current pandemic situation this plan will not be accomplish within this year.

WP2 dealing with behavior of material in frequency domain is currently at the step of validation of coded in MATLAB analytical relation between stress and strain (WT2). WT3 of WP2, related to implementation of temperature-dependence is already carried out.