Experiment-Driven Gaussian Process Surrogate Modeling and Bayesian Optimization for Multi-Objective Injection...

Area of Science:

• Materials Science
• Manufacturing Engineering
• Data Science

Background:

• Traditional injection molding (IM) optimization relies heavily on simulations, which fail to account for real-world machine variability and process noise.
• This limitation hinders the achievement of optimal part quality and manufacturing efficiency.

Purpose of the Study:

• To develop and validate an experiment-driven machine learning framework for multi-objective optimization of IM process parameters.
• To identify key parameters influencing volumetric shrinkage, warpage, cycle time, and part weight.

Main Methods:

• Conducted physical experiments using a face-centered central composite design on an industrial IM machine with high-density polyethylene.
• Benchmarked Gaussian process regression against other ML algorithms, identifying it as the best surrogate model for most quality metrics.
• Employed constrained Bayesian optimization with progressive constraint tightening for parameter optimization.

Main Results:

• Gaussian process regression demonstrated superior performance for predicting most quality metrics, though warpage prediction remained challenging.
• Holding time was identified as the most influential parameter across multiple quality responses.
• Validated optimal parameter configurations against a held-out test set, confirming the framework's effectiveness.

Conclusions:

• An experiment-driven machine learning framework offers a viable and practical alternative to simulation-based optimization for injection molding.
• This approach contributes to the advancement of experiment-centric smart manufacturing in polymer processing.