Validation Strategy Shapes Machine Learning Accuracy in Multi-Fiber Composite Design

A research team evaluating multi-fiber hybrid natural fiber epoxy composites has found that validation strategy plays as critical a role as algorithm selection when creating reliable machine learning models. Expanding upon an earlier proof-of-concept investigation that relied on only six hybrid formulations and a restricted dataset, the authors expanded their experimental design to assess fifteen distinct formulations. The work incorporated combinations of jute fiber, coconut coir, banana fiber, and pineapple leaf fiber embedded within an epoxy matrix.

The expanded experimental matrix allowed investigators to generate 3,000 specimen-level mechanical measurements across multiple testing protocols. The physical evaluations focused on four primary mechanical properties: tensile strength, flexural strength, impact strength, and Shore D hardness. By substantially increasing the volume of physical data, the researchers sought to evaluate whether modern computational regression algorithms could reliably predict mechanical properties across different composite compositions.

To rigorously assess algorithmic capabilities, the team benchmarked nine distinct regression models under two contrasting evaluation frameworks. They first applied conventional specimen-level random-split validation, which randomly parcels out individual specimens into training and testing sets. Alongside this traditional approach, the researchers implemented formulation-level GroupKFold cross-validation, designed to evaluate how models generalize when confronted with entirely separate hybrid formulations rather than isolated specimens.

The comparative evaluation exposed substantial discrepancies between the two validation strategies. Under conventional specimen-level random-split validation, all nine evaluated regression models performed well, registering high coefficient of determination values ranging from R² = 0.87 to 0.97. However, when tested under formulation-level cross-validation, model performance diverged sharply, demonstrating that high random-split accuracy can mask underlying limitations in real-world formulation prediction.

Under formulation-level cross-validation, Gaussian Process Regression demonstrated the most dependable generalization performance, maintaining an R² value between 0.73 and 0.85 across the tested properties. Linear Regression also exhibited reliable generalization, yielding R² values ranging from 0.61 to 0.80. Conversely, tree-based ensemble models and an Artificial Neural Network demonstrated poor generalization when tested at the formulation level, revealing that their predictive success relied heavily on interpolation within the sampled compositional region rather than true extrapolation.

To understand the structural influences driving mechanical responses, the researchers performed SHAP analysis on the predictive models. This interpretability analysis revealed distinct roles for individual fiber types in determining final composite performance. Banana fiber served as the primary contributor enhancing impact strength and Shore D hardness, whereas jute fiber dominated the tensile and flexural strength outcomes within the hybrid epoxy system.

Through multi-property computational optimization, the investigators determined that a formulation designated as H-15 offered the optimum overall mechanical profile. The H-15 formulation comprises 30 weight percent jute fiber, 10 weight percent coconut coir, 10 weight percent pineapple leaf fiber, and 0 weight percent banana fiber. In experimental verification, H-15 outperformed the optimum formulation established in the earlier preliminary study across tensile strength, flexural strength, and Shore D hardness.

A learning-curve analysis conducted across the dataset confirmed that expanding the experimental space from 6 formulations to 15 formulations significantly enhanced the generalization capacity of the models. The authors noted that model generalization remains confined to interpolation within the sampled compositional space rather than true extrapolation. Overall, the findings confirm that establishing proper formulation-level cross-validation protocols is just as crucial as algorithmic choice for designing natural fiber composites with machine learning.

Subject of Research: Machine learning optimization and validation strategies for hybrid natural fiber epoxy composites

Article Title: Advanced machine learning for predictive design of multi fiber hybrid natural fiber epoxy composites

Article References: Sundaravadivel, P., Sathees Kumar, S., Abinaya, G., & Pati, P. R. (2026). Advanced machine learning for predictive design of multi fiber hybrid natural fiber epoxy composites. Discover Applied Sciences. https://doi.org/10.1007/s42452-026-09779-4

Image Credits: AI Generated

DOI: 10.1007/s42452-026-09779-4

Keywords: machine learning, natural fiber composites, epoxy composites, jute fiber, coconut coir, banana fiber, pineapple leaf fiber, Gaussian Process Regression, GroupKFold cross-validation, SHAP analysis, tensile strength, flexural strength, impact strength, Shore D hardness, multi-property optimization, composite materials, regression algorithms