Advanced machine learning for predictive design of multi fiber hybrid natural fiber epoxy composites
This study addresses the limitation of our previous proof-of-concept work, which employed only six hybrid formulations and a small dataset for machine learning prediction. The experimental design was expanded to 15 distinct jute (JF)/coconut coir (CC)/banana fiber (BaF)/pineapple leaf fiber (PALF) hybrid epoxy formulations, generating 3000 specimen-level mechanical measurements (tensile, flexural, impact, and Shore D hardness). Nine regression algorithms were benchmarked using both conventional specimen-level random-split validation and formulation-level GroupKFold cross-validation. Although all models achieved high accuracy under random-split validation (R² = 0.87–0.97), formulation-level validation revealed that Gaussian Process Regression (R² = 0.73–0.85) and Linear Regression (R² = 0.61–0.80) generalized most reliably, whereas tree-based ensemble models and the Artificial Neural Network exhibited poor formulation-level generalization (interpolation within the sampled compositional region, not true extrapolation; see Sect. 4, Limitations). SHAP analysis identified banana fiber as the primary contributor to impact strength and hardness, while jute fiber dominated tensile and flexural strength. Multi-property optimization identified H-15 (30 wt% JF, 10 wt% CC, 10 wt% PALF, 0 wt% BaF) as the optimal formulation, outperforming the previously reported optimum in tensile strength, flexural strength, and hardness. Furthermore, learning-curve analysis confirmed that increasing the number of formulations from 6 to 15 significantly improved model generalization. Overall, the results demonstrate that validation strategy is as important as algorithm selection for developing reliable machine learning models for hybrid natural fiber composite design.
Authors
- Pravat Ranjan Pati (ORCID: https://orcid.org/0000-0002-8591-0219)
- G. Abinaya (ORCID: https://orcid.org/0000-0002-1907-0733)
- S. Sathees Kumar (ORCID: https://orcid.org/0000-0002-1294-8363)
- P. Sundaravadivel
Institutions
- Graphic Era University (IN)
- Saveetha University (IN)
Publication Details
- Journal
- Discover Applied Sciences
- Published
- 2026-11-12
- DOI
- https://doi.org/10.1007/s42452-026-09779-4
- Primary Topic
- Natural Fiber Reinforced Composites
- Type
- article
- Field-Weighted Citation Impact
- 0.00