Efficient Prediction of Fatigue Damage Analysis of Carbon Fiber Composites Using Multi-Timescale Analysis and Machine Learning Authors: Satoru Yoshimori, Jun Koyanagi, Ryosuke Matsuzaki Published: 2024 ABSTRACT Carbon fiber reinforced plastic (CFRP) possesses numerous advantages, such as a light weight and high strength; however, its complex damage mechanisms make the evaluation of fatigue damage particularly challenging. Therefore, this study proposed and demonstrated an entropy-based damage evaluation model for CFRP that leverages the entropy derived from heat capacity measurements and does not require knowledge of the loading history. This entropy-based fatigue degradation model, though accurate, is computationally intensive and impractical for high-cycle analysis. To address this, we reduce computational cost through multi-timescale analysis, replacing cyclic loading with constant displacement loading. Characteristic variables are optimized using the machine learning model LightGBM and the response surface method (RSM), with LightGBM achieving a 75% lower root mean squared error than RSM by increasing features from 3 to 21. This approach cuts analysis time by over 90% while retaining predictive accuracy, showing that LightGBM outperforms RSM and that multi-timescale analysis effectively reduces computational demands. 1. Introduction Carbon fiber reinforced plastic (CFRP) is used in various fields, including the aircraft [1,2], automotive [3], and medical industries [4,5] because of its advantages in lightweight properties, high strength, and high rigidity. Therefore, it is essential to establish the long-term reliability of CFRP, which is expected to be used in various fields in the future. However, fatigue damage of CFRP in the long-life region has not been sufficiently studied yet. The fracture morphology of CFRP laminates in fatigue is complex [6], unlike that of metallic materials, and the evaluation of damage behavior in terms of fracture mechanics remains insufficient. The fracture of CFRP laminates can be broadly classified into matrix cracks [7,8], which occur at the interface between the matrix and fibers without fiber fracture, transverse cracks [9,10], delamination [11] and fiber fracture [12]. In general, damage propagation in CFRP laminates under fatigue loading depends on the laminate composition. Afterwards, the stress concentration at the tip of the matrix crack causes matrix cracks in adjacent layers, transverse cracks, and delamination, which propagate into the laminate. Finally, fiber fracture leads to the failure of the CFRP laminate [6,13]. Components made from CFRP materials are often subjected to the coupled effects of fatigue load and service environment, including temperature- and humidity-induced corrosion. These adverse environmental effects can accelerate fatigue damage, such as crack formation and propagation [14,15,16]. Thus, it is very difficult to evaluate the fracture mechanics of interacting damage modes. Therefore, resolving these issues is necessary to establish the long-term reliability of CFRP laminates. Quantitative measurement of damage in CFRP laminates is challenging due to the complexity of failure modes in CFRP fatigue. In addition, it is known that there is a time dependence of damage in the polymeric material, which is the resin portion where the damage begins [17,18]. Therefore, we propose to use the entropy [19,20] generated by the damage of the resin to evaluate the lifetime of the material [21,22]. Entropy can be calculated in two ways. The first is the mechanical method, where mechanical entropy is derived from by dissipated energy. Dissipated energy is the thermal energy generated by the resistance force due to viscosity or friction, which is determined by the stress-strain history, and is a value that can be used to evaluate material damage [23]. The results of fracture simulations performed by Takase et al. [21] confirm that different combined stress states will result in failure at similar values of mechanical entropy. Therefore, mechanical entropy can be used to predict the lifetime of a material under various loading. Next, the second method is the thermodynamic method. Thermal entropy, using the thermodynamic method is calculated based on the second law of thermodynamics [24]. Thermal entropy can be calculated by measuring heat capacity using differential scanning calorimetry (DSC) [25,26] or the lock-in thermography [27,28,29] even if the loading history is unknown. Molecular dynamics simulation and experimental results by Sakai et al. [30] confirmed that mechanical and thermal entropy have similar values when the same amount of strain is applied [30,31]. Therefore, the damage behavior of CFRP with unknown loading history can be evaluated by comparing the mechanical entropy calculated by cyclic loading simulation with that calculated by measuring the heat capacity. Koyanagi et al. [32,33,34] developed a fatigue degradation model for CFRP to estimate the remaining life and residual strength of CFRP through an analytical rather than an experimental approach. The proposed model provides a practical approach that determines the strength degradation of CFRP using the entropy damage law; this approach is capable of considering not only single-cycle loading but also various complex loading histories [34]. Indeed, the proposed model enables the estimation of residual strength degradation using only the experimental measurement of entropy without requiring any knowledge of the fatigue history. However, this model has high computational cost and has not been able to analyze the number of cyclic load in accordance with the experiment (example: 100,000 cycles). Therefore, it is necessary to greatly reduce the computational cost of the analytical model. In this study, we propose reducing computational cost by replacing cyclic loading with constant displacement loading using multi-time scale analysis [35]. In this approach, the characteristic variables are created to be consistent before and after the replacement, but it is difficult to determine because of many variables. Therefore, we propose to use the response surface method (RSM) [36,37,38] and the machine learning model LightGBM [39,40,41] to model the characteristic variables. To reduce the computational cost of the analytical model, the entropy of the resin part of the CFRP was calculated using Abaqus CAE [42] under the same time cyclic loading and constant displacement loading. Next, prediction models of the characteristic variables were created using this data, response surface method and LightGBM. Finally, the prediction accuracy of the response surface method and LightGBM was evaluated by predicting entropy under cyclic loading using the prediction model of the characteristic variables and comparing this value with entropy value under cyclic loading calculated using Abaqus CAE. 2. Multi-Timescale Analysis and Prediction Modeling Methods 2.1. Multi-Timescale Analysis Multi-timescale analysis is a method of replacing changes in displacement with constant displacement and characteristic variables on small oscillating timescales, as shown in Figure 1. The characteristic variables are added to the oscillating displacement under constant displacement conditions and modifies material constants to ensure the coherence of the entire system. Using this method, Takamura et al. [35] replaced ultrasonic vibrations with constant stress and characteristic variables, and succeeded in connecting phenomena of the order of 10−4 with those of the order of 100 in their numerical simulations [35]. In this way, the analysis of cyclic loading, which require many calculations with detailed frequency, can be performed with fewer calculations by replacing cyclic loading with constant displacement loading and characteristic variables, as shown Figure 1. As an example, this method can reduce the number of calculation steps from 10 steps per cycle × 100 cycles with a cyclic loading to 1/1000 by replacing it with a single time increment with a constant displacement loading. Given the above, we use multi-timescale analysis to reduce the computational cost of the analytical model. Based on the idea of multi-timescale analysis, we model the characteristic variables and use the model to predict entropy under cyclic loading while reducing computational cost as shown Figure 1. These aspects will be explained in later sections. 2.2. LightGBM and Response Surface Method Next, the modeling of the characteristic variables are described. Determining the characteristic variables in a multi-timescale analysis is challenging because many variables, such as amplitude and frequency, influence the entropy increase. Therefore, we propose using LightGBM and response surface method to create prediction models of the characteristic variables. LightGBM [39] is a decision tree-based gradient boosting algorithm and a representative machine learning model. As shown in the Figure 2a, LightGBM models the objective function by repeatedly calculating predictions using decision trees and learning residuals, which are the differences between the predictions and the objective function. The response surface method [36], on the other hand, generates a response surface as shown in Figure 2b based on response values obtained from experiments and analyses, and creates a regression model as shown in Equation (1). Here, y is the response, xi is the explanatory variables, n is the number of explanatory variables, β0, βi, βii, and βij are the regression coefficients. In this study, we used these two methods to predict entropy under cyclic loading and compared their prediction accuracy. (1)y=β0+∑i=1nβixi+∑i=1nβiixi2+∑i