پيش‌بيني رفتار كششي فيلامنت‌هاي كامپوزيتي تقويت‌شده با الياف كوتاه با استفاده از مدل ميكرومكانيك و يادگيري ماشين

Prediction of tensile behavior of short fiber-reinforced filament composites using micromechanics models an‎d machine learning

مهندسي مكانيك

In this study, to investigate an‎d optimize ABS–glass composite filaments for materials engineering an‎d additive manufacturing applications, the filaments were first produced via material extrusion. Following fabrication, the specimens were subjected to a series of experimental tests, including uniaxial tensile testing, furnace treatment, an‎d micro-CT imaging. The results of these experiments, particularly the micro-CT data, served as the main basis for detailed analyses an‎d micromechanical modeling using Digimat software. The concurrent use of multiple experimental methods not only enabled the determination of key mechanical properties but also allowed for the eva‎luation of the accuracy an‎d validity of the initial manufacturing assumptions. Subsequently, micromechanical modeling of the tensile behavior of short-fiber-reinforced composite filaments was performed using Digimat-MF. To ensure the reliability of the modeling process, the simulated results were compared with the experimental data. After successful validation, an extensive parametric study was conducted to examine the effects of critical factors, including fiber volume fraction, void volume fraction, fiber aspect ratio, an‎d tensile loading direction. Based on the modeling results, a database comprising 100 stress–strain curves was developed an‎d used for statistical analysis as well as for constructing an intelligent predictive framework. In this framework, five machine learning algorithms, Decision Tree (DT), K-Nearest Neighbors (KNN), Elastic Net (EN), Ran‎dom Forest (RF), an‎d Gaussian Process Regression (GPR), were compared for the simultaneous prediction of elastic modulus an‎d ultimate tensile strength. In addition, two deep learning algorithms, Simple Recurrent Neural Network (Simple RNN) an‎d Long Short-Term Memory (LSTM), were employed to predict the entire stress–strain response. The results revealed that micro-CT testing accurately captured the composite microstructure an‎d provided the most reliable data for micromechanical modeling. The discrepancy between the micromechanical modeling an‎d experimental results was less than 10%, confirming the high accuracy of the modeling process. In the parametric study, increasing the fiber volume fraction from 1% to 15% led to a 31.785% increase in elastic modulus an‎d a 15.968% increase in ultimate tensile strength. Likewise, increasing the fiber aspect ratio from 1 to 10 enhanced the elastic modulus by 5.449% an‎d the ultimate tensile strength by 6.268%. Conversely, increasing the void volume fraction reduced both properties, while varying the loading direction had negligible influence due to the material’s isotropy. Data analysis further indicated that fiber volume fraction was the most influential factor affecting elastic modulus an‎d ultimate tensile strength, with correlation coefficients of 0.9 an‎d 0.84, respectively. Among the machine learning models, GPR achieved the highest prediction accuracy (99.997%) for simultaneous prediction of elastic modulus an‎d tensile strength. Similarly, LSTM yielded the highest accuracy (99.995%) in predicting the complete stress–strain curve. Overall, the results demonstrate that combining high-precision experimental methods with micromechanical modeling in Digimat-MF provides an accurate an‎d robust framework for predicting the mechanical behavior of extrusion-based ABS–glass composite filaments.

5