Machine Learning-Assisted Design of Multilayer Thermoplastic Composites: Robust Neural Network Prediction and Feature Importance Analysis

Abstract

Multilayer thermoplastic composites offer sustainable alternatives to traditional thermoset and metal materials. However, their design is inherently complex, involving numerous interdependent parameters that render conventional processes both expensive and time-consuming. While machine learning-assisted methods provide a potential solution, they typically require large datasets that can be costly to obtain. This study explores a robust neural network, specifically, an Advanced Multilayer Perceptron (AdvMLP) Regressor, to predict the peel strength of multilayer thermoplastic composites. Through architectural enhancements, the AdvMLP is effectively trained on a limited yet authentic manufacturing dataset, yielding robust predictions validated by benchmark metrics and k-fold cross-validation. The model captures the intricate interplay between manufacturing processes and composite properties, enabling comprehensive feature importance analysis and dimensionality reduction. Overall, this study establishes a robust and generalizable machine learning-assisted methodology to guide and accelerate the design and optimization of multilayer thermoplastic composites.