Machine Learning Optimization of Alkali-treated Sisal Fiber Reinforced Polyester Composites for Mechanical and Water Resistance Performance

Abstract

Aims: The study characterizes the effects of NaOH concentration, soaking duration, and fiber-to-polyester ratio on the mechanical performance, FTIR spectral signatures, and water resistance of sisal/polyester composites, and to evaluate machine learning (ML) frameworks for predictive property modelling from these experimental parameters. Study Design:  A full-factorial experimental design (3 × 3 × 3) was adopted to systematically evaluate the effects of NaOH concentration, soaking time, and fiber-to-polyester ratio on composite performance. This design was integrated with Principal Component Analysis (PCA) for analysing FTIR spectral data and Support Vector Regression (SVR) modelling. Model validation was performed using leave-one-out cross-validation (LOOCV) to ensure robustness, particularly given the limited dataset. Methodology: Sisal fibers were treated with NaOH solutions at concentrations of 2%, 6%, and 10% (w/v) for durations of 24, 48, and 72 hours. Composite samples were fabricated using the hand lay-up method at fiber-to-polyester ratios of 20:80, 30:70, and 40:60 (w/w), resulting in 27 composite specimens alongside fiber-level tensile observations. FTIR analysis was conducted using a JASCO Model 4100 spectrometer within the range of 400–4000 cm⁻¹. Mechanical characterisation included single-fiber tensile testing (ASTM C1557), composite tensile testing (ASTM D638), flexural testing (ASTM D790), and water absorption testing (ASTM D570). Machine learning analysis was performed using the scikit-learn library with StandardScaler normalization, and permutation feature importance was computed over 50 repetitions. Results: The optimal composite performance was observed at 6% NaOH concentration, 48 hours soaking time, and a 30:70 fiber-to-polyester ratio, yielding a tensile strength of 44.003 MPa and bending strength of 50.810 MPa. The lowest water absorption (3.194%) was achieved at 6% NaOH, 72 hours, and a 20:80 ratio. PCA identified four distinct fiber treatment groups based on FTIR spectral features, indicating clear chemical differentiation due to treatment conditions. The SVR model achieved a moderate predictive performance (R² = 0.357; MAE = 14.55 MPa) under LOOCV using the fiber tensile dataset. Feature importance analysis revealed NaOH concentration as the most influential variable affecting mechanical performance. Conclusion: Mercerization at 6% NaOH / 48 h achieves optimal surface activation without structural cellulose damage. The FTIR-to-property ML prediction pathway demonstrates feasibility for non-destructive composite quality assessment. Future work should address dataset expansion and multi-objective optimization.