Data-driven Modeling of Mechanical Behaviors of Additively Manufactured Materials
Saved in:
| Title: | Data-driven Modeling of Mechanical Behaviors of Additively Manufactured Materials |
|---|---|
| Authors: | Zhang, Ziyang |
| Committee Members: | Wu, Dazhong |
| Summary: | Additive manufacturing (AM) is a revolutionary technology that greatly improves the flexibility of fabricating parts with complex structures and eliminates the cost of making molds. While AM techniques offer unique benefits over traditional manufacturing processes, it is challenging to predict the mechanical behaviors of additively manufactured parts based on design and process parameters. With recent advances in machine learning, data-driven methods have the potential to overcome such limitations. In this work, data-driven modeling frameworks were proposed to predict the tensile, flexural, and compressive behaviors of additively manufactured plastics and composites. Ensemble learning was used to predict the tensile strength of polylactic acid (PLA) with cooperative AM process parameters. A 12.97% mean absolute percentage error (MAPE) was achieved by combining lasso, support vector regression, and extreme gradient boosting in the computational framework. An enhanced ensemble learning method that combines eight different machine learning algorithms was introduced to predict the flexural strength of continuous carbon fiber and short carbon fiber reinforced nylon (CCF-SCFRN) composites with design parameters. Learned knowledge from CCF-SCFRN composites was transferred to continuous glass fiber and short carbon fiber reinforced nylon (CGF-SCFRN) composites for flexural stress-strain curve prediction using an optimal transport (OT) integrated transfer learning framework. Compared with traditional transfer learning, the OT-integrated framework improves the stress-strain curve prediction accuracy by 10.46% in terms of MAPE. The transfer learning framework was further demonstrated in predicting the compressive stress-strain curves of PLA scaffolds with both AM process and design parameters. Three cases were studied by selecting different parameters for domain transfer to validate the generalizability of the proposed framework in predicting mechanical behaviors of additively manufactured materials with limited data. |
| URL: | https://stars.library.ucf.edu/etd2020/1454 |
| Database: | OpenDissertations |
| Abstract: | Additive manufacturing (AM) is a revolutionary technology that greatly improves the flexibility of fabricating parts with complex structures and eliminates the cost of making molds. While AM techniques offer unique benefits over traditional manufacturing processes, it is challenging to predict the mechanical behaviors of additively manufactured parts based on design and process parameters. With recent advances in machine learning, data-driven methods have the potential to overcome such limitations. In this work, data-driven modeling frameworks were proposed to predict the tensile, flexural, and compressive behaviors of additively manufactured plastics and composites. Ensemble learning was used to predict the tensile strength of polylactic acid (PLA) with cooperative AM process parameters. A 12.97% mean absolute percentage error (MAPE) was achieved by combining lasso, support vector regression, and extreme gradient boosting in the computational framework. An enhanced ensemble learning method that combines eight different machine learning algorithms was introduced to predict the flexural strength of continuous carbon fiber and short carbon fiber reinforced nylon (CCF-SCFRN) composites with design parameters. Learned knowledge from CCF-SCFRN composites was transferred to continuous glass fiber and short carbon fiber reinforced nylon (CGF-SCFRN) composites for flexural stress-strain curve prediction using an optimal transport (OT) integrated transfer learning framework. Compared with traditional transfer learning, the OT-integrated framework improves the stress-strain curve prediction accuracy by 10.46% in terms of MAPE. The transfer learning framework was further demonstrated in predicting the compressive stress-strain curves of PLA scaffolds with both AM process and design parameters. Three cases were studied by selecting different parameters for domain transfer to validate the generalizability of the proposed framework in predicting mechanical behaviors of additively manufactured materials with limited data. |
|---|
