Machine Learning Applied to Composite Materials

Every engineering application requires a comprehensive investigation of different parameters in different exposure conditions to come up with an optimal yet feasible product. Modeling the complex relationships between the various governing factors is extremely strenuous and generally requires the development of a mathematical tool. This has motivated researchers to look for time saving and less expensive computational techniques. Machine learning is perceived as the next big wave of innovation. It has revolutionized the field of Material Science by showing promising results when it comes to target oriented research. In the past few years, machine learning has been used as an efficient tool for predicting the material behavior but still there are inhibitions to use the various algorithms for large scale implementation. In an attempt to provide perspective on the usage of machine learning algorithms in polymer composites, this chapter summarizes the recent studies conducted on these composite materials using machine learning along with a general overview of its multifaceted applications like prediction, optimization, uncertainty quantification and sensitivity analysis.

The products manufactured by the additive manufacturing (AM) methods have unique signatures in their microstructures due to the layer by layer manufacturing. Machine learning of microstructures of the printed sample can help in interpreting these signatures and the patterns can be used for either determining the authenticity of the product or for reverse engineering. In this work, specimens of 3D printed glass fiber reinforced polymer (GFRP) composite materials are subjected to imaging and machine learning in order to rebuild the tool path information. Since composites require significant research and development effort, the possibility of rebuilding the tool path by ML methods presents a significant vulnerability for intellectual property. The ML methods require a large training dataset and can be efficient in processing tomography datasets. Two kinds of artificial neural networks with three different algorithms are introduced in this work and their results are compared. A 3D printed GFRP specimen is imaged using a micro CT-scan and the images are processed using binarized statistical image features method for compression without compromising the microstructural information. The ML models are trained on this dataset and the results indicate that the ML is able to identify the printing tool path with accuracy.

Property variability is inevitably introduced into heterogeneous composites through manufacturing processes, service loads, or environmental factors, and such variability strongly impacts macroscopic structural performance. Uncertainty Quantification provides a framework for measuring these uncertainties and assessing their impact, increasing confidence in simulation results. Rigorous UQ can pave the way toward certification by simulation of composite materials, manufacturing processes, and aerostructures, where extensive experimental effort is currently required. This body of work explores UQ methods in computational efficiency, assessment of material variability, and investigates novel artificial intelligence and machine learning techniques to provide a more general and interpretable framework for UQ. First, a method to reduce the computational burden of the Monte Carlo (MC) method for propagation of uncertainty by adapting the Quasi Monte Carlo (QMC) method with different low discrepancy sampling techniques is considered. A considerable increase in convergence speed is shown for the QMC technique compared to MC. Furthermore, it is shown that using Convolutional Neural Network (CNNs), the parameters describing random fields modeling the spatial property variability of a material can be inferred from a few experimental tests with full-field strain measurement. The proposed methodology allows examination of the stochastic response and uncertainty quantification of additively manufactured structures, while requiring only minor experimental efforts to fully define the random fields. Once the CNNs are trained, computational expense for predicting stochastic parameters is minimal.

This chapter presents the stochastic low-velocity impact responses of a cantilever plate composed of functionally graded materials (FGM) by employing the radial basis function (RBF) model. A novel algorithm is developed to ascertain the stochastic transient impact responses experienced by the FGM plates due to the central impact of a spherical steel ball impactor. The random variabilities in the material properties and temperature are considered in the present study. A convergence study of the surrogate model is carried out in conjunction to sample sizes, and results are validated with the original Monte Carlo simulation (MCS). The percentage of error present in the constructed RBF model is also determined. The influence of different input variables such as the degree of stochasticity, power-law exponent, temperature, oblique impact angle, and initial velocity of the impactor is considered for mapping the stochastic transient low-velocity impact parameters such as maximum contact force (CF), maximum plate displacement (PD), and maximum impactor displacement (ID). The statistical results illustrate that all input parameters have significant effects on the stochastic impact responses of cantilever FGM plates.

Currently, the challenge in front of researchers is to discover new novel material with superior properties as per the demand of the society with a vast range of applications. With evaluation in material characterization techniques large amounts of material data are obtained through experiments and simulations. Even in some cases theoretical concepts cannot be applicable to these data. With increase in material data, application of machine learning and data analytics come into play. Application of machine learning is applicable in various fields such as material properties, analyzing complex reactions, inorganic chemistry, understanding crystal structure, in the design of experiments, etc. Through this article our focus is towards application of machine learning in the field of material characterization techniques in determining the mechanical properties of materials. In this chapter, a brief review of application of machine learning in the field of characterization of the mechanical properties such as tensile strength, fatigue behavior and visco-elastic study have been done.

Composite materials have found their wide variety of applications in innumerable sectors such as aerospace, automotive and marine sectors. Lightweight and high specific mechanical properties make composite materials a successor to those conventional metal alloys. Another benefit of composite materials is that their mechanical properties can be designed to meet the criteria for certain engineering applications. However, the mechanical properties of composite materials are governed by several factors such as fibre content, fibre alignment, fibre-matrix compatibility, etc. Therefore, it is essential to identify the optimum parameters in order to develop composite materials at their exemplary mechanical performance. Today, several advanced techniques have been developed to accurately predict the mechanical properties of composite materials without conducting experimental works. With these advanced techniques to predict the mechanical performance of composite materials, costs and time for the material preparation and experimental works can be drastically reduced. This chapter intends to discuss the background of lightweight composite materials. Additionally, insight into the prediction of the mechanical properties of lightweight composites via machine learning is included in this chapter to understand the non-destructive testing methods for composite materials.

In the present article, we explored the ballistic behaviour of bilayer graphene (BLG) by performing a series of molecular dynamics (MD) simulations. The computationally expensive nature of large scale MD simulations frequently hinders a thorough understanding of material characterization. To mitigate this lacuna we demonstrated the successful integration of MD simulation with the artificial neural network (ANN). In this regard, the considered input parameter [impact velocity (V_i)] is perturbed in the range of 1–7 km/s using the Monte Carlo sampling technique to construct the sample space with 128 instances. The BLG (size 200 Å × 200 Å) is impacted by a spherical diamond projectile (diameter 25 Å) in a series of MD simulations of high-velocity impact with varied impact velocities. As a response, the residual velocity of the projectile (V_r) and specific penetration energy (\(E_{p}^{*}\)) of the BLG are determined for each instance. The deterministic responses revealed that with the increase in the impact velocity the V_r and \(E_{p}^{*}\) values increases. Besides the numerical responses, the post-impact behaviour of BLG is also classified into four different stages viz. R, PP₁, PP₂ and CP, based on the extent of damage to the BLG and the post-impact trajectory of the projectile. The dataset generated with the MCS based MD simulation is further used to construct the ANN based regression and classification model. In this manner, the current article proposed a framework to accelerate the nanoscale material characterization by augmenting the ANN with MD simulations.

The steel girder bridge load ratings are typically based on 1-dimensional (1D) AASHTO line girder analysis which can often be conservative for effective bridge management decisions within constrained fiscal resources. AASHTO allows more rigorous and refined bridge analysis methods, but these approaches can provide an uncertain investment return. Recent studies have demonstrated that machine learning-based methods, such as artificial neural networks (ANNs), can be effectively used as a predictive tool to complement and anticipate the likely outcome of more rigorous bridge load rating analyses. The present study quantifies the sensitivity of input governing parameters for two ANN-based bridge loading rating models: (1) single-best-network and (2) committee networks (CN). The variance of prediction-model outputs was used to measure the sensitivity of network input-parameters when perturbed around their mean values. For illustration as a case study, refined moment-based load ratings were obtained on 254 steel girder bridges using 3D-finite element analysis (FEA) and then used as datasets for input parameter sensitivity analysis of ANN-based models. Among the parameters of study for steel girder bridges (i.e., structural and geometric bridge characteristics), the spacing of girders in the CN model and the barrier-edge distance parameter in the single-best-network model showed the most influence and sensitivity (about 46% in this study) on the moment-based load rating outputs.

Concrete-filled double-skin tubular (CFDST) columns are optimized for a high strength-to-weight ratio by having their concrete core confined between inner and outer steel tubes. The confined concrete behavior in these composite columns is affected by the shape of the inner and outer steel tubes. A new hybrid approach using genetic algorithm (GA)-optimized artificial neural networks (ANNs) is proposed in this study to estimate the axial load capacity of CFDST columns for multiple combinations of square and circular steel tubes, i.e., circle-circle (CC), circle-square (CS), square-square (SS) and square-circle (SC) cross-sections. The present study used a total dataset of 171-CFDST columns (i.e., 51 of CC, 43 of CS, 38 of SC and 39 of SS shapes) for demonstration. The axial load capacity of CFDST columns was obtained using calibrated nonlinear finite element analyses. For the training of ANNs, a design set of 100 CFDST columns was used from among hypothetical set to map geometric and material properties to their ultimate axial load capacity. The remaining 71 columns (combined hypothetical and actual tested specimens) were used to test and check the generalization ability of the ANN-based prediction model. The network parameters of ANNs were reoptimized with GA to reduce the maximum absolute error on the testing set columns from 26 to 14%. Thus, the hybrid GA-optimized ANNs can more accurately predict the ultimate axial load capacity of CFDST columns of multiple shapes (about 14 and 7%, respectively, the maximum and mean absolute errors in this study) than traditional ANNs.