Interpretable Prediction and Analysis of PVA Hydrogel Mechanical Behavior Using Machine Learning.

Area of Science:

• Materials Science
• Polymer Science
• Biomaterials Engineering

Background:

• Polyvinyl alcohol (PVA) hydrogels are promising for biomedical applications due to biocompatibility and tunable properties.
• Rational design of PVA hydrogels is hindered by complex structure-property relationships.
• Existing machine learning models often lack mechanistic interpretability.

Purpose of the Study:

• To develop an interpretable machine learning framework for predicting PVA hydrogel tensile strain properties.
• To gain mechanistic insights into the structure-property relationships governing PVA hydrogel performance.
• To enable rational design strategies for advanced PVA hydrogels.

Main Methods:

• Compiled a dataset of 350 data points from a systematic literature review.
• Employed an XGBoost model optimized using Optuna for performance prediction.
• Utilized SHAP analysis to interpret model predictions and identify key influential parameters.

Main Results:

• Achieved high predictive accuracy with R² values of 0.964 (training) and 0.801 (testing).
• Identified PVA molecular weight (SHAP importance: 84.94) as the dominant factor, followed by degree of hydrolysis and cross-linking.
• Elucidated complex non-linear relationships and reinforcement mechanisms through interpretability analysis.

Conclusions:

• The interpretable machine learning framework successfully predicts PVA hydrogel properties and provides mechanistic understanding.
• PVA molecular weight, degree of hydrolysis, and cross-linking are critical for mechanical performance.
• This approach facilitates the rational design of next-generation multifunctional PVA hydrogels.