Predicting flexural properties of 3D-printed composites: A small dataset analysis using multiple machine learning models

Fused Deposition Modelling (FDM) is a common Additive Manufacturing (AM) technique for processing polymers and composites. Infill density is one of the crucial parameters affecting the mechanical properties of the processed materials/structures. In material science, the availability of large experimental datasets acquired under the same conditions is a costly and time-consuming practice. To address this issue, the present work focuses on deploying five machine learning models including Linear Regression (LR), Polynomial Regression (PR), Gaussian Process Regression (GPR), Gradient Boosting (GB), and Random Forest (RF) to test their prediction capabilities with a small experimental flexural load-extension dataset of Short Carbon Fiber Reinforced PLA (CF-PLA) consisting of only 550 records. Experimental results yielded a direct relation between the infill density and flexural performance of CF-PLA. An increase of 308 %, 280 %, and 94 % was observed for maximum flexural load, flexural strength, and flexural modulus, respectively. Following this, to access and compare the performance of each ML model in forecasting these improvements, Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and R-squared (R-2) metrics were used. The results revealed that the Random Forest (RF) algorithm successfully predicted material behavior with an MAE, RMSE, and R-2 of only 0.64, 1.17, and 0.99, respectively. Whereas, LR was found to be least effective in capturing the complex behavior of the composite with an MAE, RMSE, and R-2 of 6.94, 9.71, and 0.82, respectively. Thereby indicating the efficacy of the RF algorithm for even smaller data sets.