Machine Learning Enhances Composite Impact Predictions

By Soham NandiReviewed by Susha Cheriyedath, M.Sc.Jul 23 2024

In an article published in the journal Results in Engineering, researchers explored predicting the impact behavior of fabric-layered composites, specifically Kevlar and carbon fabrics, using machine learning models.

Low-velocity impact (LVI) tests were conducted, and the resulting data trained models to predict impact force, absorbed energy, and displacement. The study found varying impacts based on fabric type and highlighted the accuracy of different machine learning models in predicting impact behaviors and failure modes.

Background

Fabric laminated composites, particularly those using carbon and Kevlar fibers, are crucial in safety and surveillance due to their high specific strength, lightweight nature, stiffness, and durability. These composites replace metal components in various applications but are vulnerable to complex behaviors, especially impact resistance.

Previous research focused on experimental and numerical analyses to understand these behaviors, with studies on multilayer fabric composites revealing that parameters like hybridization, thickness, and stacking sequence significantly influence impact performance. Despite advancements, gaps remain regarding the precise impact of layer structures and hybrid configurations.

This paper addressed these gaps by conducting LVI tests on bidirectional carbon and Kevlar fiber laminates with varying parameters. Using data from 33 different setups, the study trained four machine-learning models to predict impact force, absorbed energy, and displacement.

The study demonstrated that machine learning can effectively predict impact behaviors, providing higher accuracy and insights into failure modes, thus enhancing the understanding and performance of laminated composites under LVIs.

Methods 

The materials used included aramid fibers with specific weights and densities, produced by Geun Young Industrial Company Limited, and epoxy resin adhesives manufactured by 3M. Laminated composites were prepared using a hand layup method, involving the stacking of cut fabric layers with mixed epoxy resin. The composite layers included pure carbon, pure Kevlar, and hybrid combinations, with thicknesses ranging from 0.25 mm to 1.04 mm, and were cured under compression.

To evaluate the impact behavior, LVI tests were conducted using a drop-weight impact machine with a cone-shaped impactor. The tests varied impact energies and heights, converting the impactor's potential energy to kinetic energy. Each sample was tested three times to ensure accuracy, resulting in 99 samples being reduced to 33 by averaging the results. The impact data, including force, velocity, displacement, and absorbed energy, were recorded using specialized software.

Four machine learning models—linear regression, polynomial regression, support vector regression (SVR), and multi-layer perceptron (MLP)—were employed to predict the impact properties. The authors outlined the theoretical basis of these models and their implementation using the scikit-learn Python package.

Data splitting procedures involved an 80:20 ratio for training and testing sets, and hyperparameters for each model were optimized to improve prediction accuracy. Model performance was assessed using statistical measures such as coefficient of determination (R-squared), mean squared error (MSE), and mean absolute error (MAE). Feature importance and partial dependence plots were used to interpret the models, providing valuable insights into the complex impact behavior of these materials under low-velocity conditions.

Results and Discussion

The impact behavior of various fiber-laminated composites was tested across different energy levels, revealing distinct performance differences. Carbon and Kevlar composites showed increased peak force with higher energy levels, with carbon composites exhibiting greater strength and stiffness. Mixed laminates like carbon-Kevlar-carbon-Kevlar (CKCK) and Kevlar-carbon-Kevlar-carbon (KCKC) displayed reduced peak force due to material combination effects.

Displacement tests indicated uniform displacement in carbon and Kevlar composites, with less displacement in mixed laminates. Energy absorption tests showed that carbon and Kevlar composites absorbed more energy at higher levels, with KCKC stacking absorbing the most.

Machine learning models, including linear regression, polynomial regression, SVR, and MLP, were evaluated for predicting impact behavior. While MLP performed well in force prediction, SVR excelled in displacement and absorbed energy prediction, suggesting potential overfitting in some models.

Experimental tests highlighted failure modes such as delamination, matrix cracking, fiber breakage, and interlaminar shear failure. These insights emphasized the need for advanced characterization techniques and predictive modeling to enhance laminated composite materials' impact resistance and structural integrity.

Conclusion

In conclusion, the researchers investigated the impact behavior of carbon and Kevlar fabric laminated composites using LVI tests and machine learning models. Key findings included the superior impact resistance of carbon and Kevlar composites, with mixed configurations like KCKC showing high energy absorption.

Machine learning models, particularly polynomial regression and SVR, effectively predicted impact force, displacement, and absorbed energy with high accuracy. The authors highlighted the importance of advanced characterization and predictive modeling in enhancing the performance and understanding of laminated composites under impact conditions. Future work aims to expand the experimental database for improved machine learning model development.