Transfer learning-guided process-structure–property mapping in dual-material fused filament fabrication

A new framework of a deep-level transfer learning model was designed for accurately making predictions of the mechanical properties of Dual Material Composites using the Fused Filament Fabrication process. The critical problem of precisely forecasting the properties of Polyethylene Terephthalate Glycol-Thermoplastic Polyurethane (PETG-TPU) and Polyethylene Terephthalate Glycol-Nylon (PETG-Nylon) composite materials in a data-scarce setting was effectively dealt with. The Dual Material Composites were deliberately created using a combination of different infill patterns (hexagonal, linear, grid, and concentric) and infill densities (30–80%) for a total of six different material distribution patterns. The three-point bending tests were conducted according to the ISO 178 standards. The new deep learning model combined convolutional neural networks for spatial feature learning, long short-term memory networks for learning temporal relationships, as well as self-attention mechanisms for adaptive feature weighting. By using a transfer learning approach for training the model, extensive PETG-Nylon datasets were used for training. The pre-trained PETG-Nylon model was fine-tuned using the PETG-TPU datasets. The proposed model performed very well in predictive tasks with a root-mean-square error of less than 2.96. The mean absolute percentage error was less than 4.03%. The value of the coefficient of determination was found to be 0.987. The degree of similarity between the predicted and experimental stress–strain behavior was found to be very satisfactory. The new approach resulted in a remarkably improved error margin when compared to a baseline model trained using a conventional learning strategy. The baseline models were found to possess a root mean square error of up to 12.53 units. The mean absolute percentage error was above 13% when using the baseline models. The critical finding of the study was that Dual Material Composites consisting of linear infills with 70% infill density were found to possess the maximum value of Flexural Strength. The new PETG-TPU prediction framework using a design-holdout strategy for a deep-level PETG-TPU learning model thus aided in efficiently determining the mechanical properties of unfamiliar PETG-TPU Dual Material Composite designs.