Data-Driven Design of Thermoplastic Composites with Tailored Compliance

Continued advances in manufacturing processes have rapidly increased the complexity of composite parts that can be manufactured, consequently increasing the number of variables that must be considered during design. Traditional computational methods struggle with time efficiency and complex relations in large design spaces, especially for composite materials with intricate, multivariate, and nonlinear process-structure-property connections. Among computational tools, machine learning methods excel at interpreting complex relationships and efficiently generating designs for target properties. In this study, we systematically built and deployed machine learning models to explore the design space of a layered composite material. We modeled two grades of thermoplastics and generated a data set of the compliance of different layered composite geometries using finite-element simulations. A random forest model utilizing a token-counting featurization scheme was selected based on its exceptional performance. It was used to perform a detailed feature importance analysis and then a series of design tasks. We show that this method can reliably obtain single-objective and multiobjective designs. This work demonstrates the feasibility of a simple data-driven approach to designing composite parts with many design variables and highly nonlinear mechanical behavior.