Hybrid Experimental–Machine Learning Study on the Mechanical Behavior of Polymer Composite Structures Fabricated via FDM

Ulkir O., ERSOY S.

Polymers, cilt.17, sa.15, 2025 (SCI-Expanded) identifier identifier

  • Yayın Türü: Makale / Tam Makale
  • Cilt numarası: 17 Sayı: 15
  • Basım Tarihi: 2025
  • Doi Numarası: 10.3390/polym17152012
  • Dergi Adı: Polymers
  • Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Academic Search Premier, Aerospace Database, Chemical Abstracts Core, Communication Abstracts, Compendex, Food Science & Technology Abstracts, INSPEC, Metadex, Directory of Open Access Journals, Civil Engineering Abstracts
  • Anahtar Kelimeler: additive manufacturing, fused deposition modeling, machine learning, mechanical behavior, polymer composite
  • Marmara Üniversitesi Adresli: Evet

Özet

This study explores the mechanical behavior of polymer and composite specimens fabricated using fused deposition modeling (FDM), focusing on three material configurations: acrylonitrile butadiene styrene (ABS), carbon fiber-reinforced polyphthalamide (PPA/Cf), and a sandwich-structured composite. A systematic experimental plan was developed using the Box–Behnken design (BBD) to investigate the effects of material type (MT), infill pattern (IP), and printing direction (PD) on tensile and flexural strength. Experimental results showed that the PPA/Cf material with a “Cross” IP printed “Flat” yielded the highest mechanical performance, achieving a tensile strength of 75.8 MPa and a flexural strength of 102.3 MPa. In contrast, the lowest values were observed in ABS parts with a “Grid” pattern and “Upright” orientation, recording 37.8 MPa tensile and 49.5 MPa flexural strength. Analysis of variance (ANOVA) results confirmed that all three factors significantly influenced both outputs (p < 0.001), with MT being the most dominant factor. Machine learning (ML) algorithms, Bayesian linear regression (BLR), and Gaussian process regression (GPR) were employed to predict mechanical performance. GPR achieved the best overall accuracy with R2 = 0.9935 and MAPE = 11.14% for tensile strength and R2 = 0.9925 and MAPE = 12.96% for flexural strength. Comparatively, the traditional BBD yielded slightly lower performance with MAPE = 13.02% and R2 = 0.9895 for tensile strength. Validation tests conducted on three unseen configurations clearly demonstrated the generalization capability of the models. Based on actual vs. predicted values, the GPR yielded the lowest average prediction errors, with MAPE values of 0.54% for tensile and 0.45% for flexural strength. In comparison, BLR achieved 0.79% and 0.60%, while BBD showed significantly higher errors at 1.76% and 1.32%, respectively.