Optimizing bio-hybrid composites for impact resistance using machine learning

This study pioneers an integrated approach combining experimental analysis and machine learning (ML) predictions to assess the low velocity impact (LVI) response of synthetic/natural bio-hybrid fiber-reinforced polymer (HFRP) composite materials. Five different stacking sequences of carbon/flax bio-HFRP were tested for LVI with impact energies from 15 to 90 J, and data such as peak impact force, damage area, and damage extension were recorded. Symmetric configuration with consistent dispersal of natural flax fibers across laminate demonstrated improved impact resistance. Furthermore, six ML algorithms were used: decision tree (DT), random forest, deep neural network with Adam optimizer (DNN-Adam), DNN with stochastic gradient (SGD) optimizer (DNN-SGD), recurrent neural network (RNN) with Adam optimizer (RNN-Adam), and RNN with SGD optimizer (RNN-SGD). Model performance was evaluated using coefficient of determination (R2), mean square error (MSE), and mean absolute error (MAE). The DT ML model achieved best performance in predicting peak impact force having maximum depth count of 8 and leaf nodes count of 28. For damage area, again, DT model with maximum depth count of 6 and leaf nodes count of 23 exhibited better performance. On the other hand, for damage extension, the RNN-SGD model, having four hidden layers and 70 neurons, outperformed other ML models. Among the investigated parameters, the highest correlation (R2 = 0.9987 for training and 0.9922 for test datasets) and lowest errors (MSE = 0.0294 and MAE = 0.1344) were achieved for predicting damage extension. This study is the first to apply advanced ML techniques to predict mechanical responses such as peak impact force, damage area, and damage extension in carbon/flax bio-HFRP composites under LVI conditions, enhancing accuracy and reducing the testing, thereby optimizing resources and significantly minimizing time.