Abstract
Heated tool butt welding is a method often used for joining thermoplastics, especially when the components are made out of different materials. The quality of the connection between the components crucially depends on a suitable choice of the parameters of the welding process, such as heating time, temperature, and the precise way how the parts are then welded. Moreover, when different materials are to be joined, the parameter values need to be tailored to the specifics of the respective material. To this end, in this paper, three approaches to tailor the parameter values to optimize the quality of the connection are compared: a heuristic by Potente, statistical experimental design, and Bayesian optimization. With the suitability for practice in mind, a series of experiments are carried out with these approaches, and their capabilities of proposing well-performing parameter values are investigated. As a result, Bayesian optimization is found to yield peak performance, but the costs for optimization are substantial. In contrast, the Potente heuristic does not require any experimentation and recommends parameter values with competitive quality.
Introduction
In manufacturing, particularly within the automotive sector, joining dissimilar thermoplastics—such as transparent PMMA for taillights and opaque ABS for housings—is a critical challenge. Heated tool butt welding is the industry standard, yet its efficacy relies on optimizing five interdependent parameters: heating temperatures and times for each component, and the joining path or force. Traditionally, this optimization is a manual "trial and error" process. This study evaluates whether modern statistical or machine learning (ML) methods can surpass traditional engineering heuristics in weld quality and efficiency.
Methods
The research compares three distinct methodologies for parameter selection:
- The Potente Heuristic: An established engineering approach using formulas based on material properties like glass transition temperature ($T_G$) and thickness. It requires virtually no physical testing to generate initial parameters.
- Statistical Experimental Design (DoE): A classical statistical method using partial factorial test plans to systematically explore the parameter space and model variable interactions.
- Bayesian Optimization (BO): An ML approach for black-box optimization. Utilizing the SMAC framework, it builds surrogate models (Random Forests or Gaussian Processes) to predict optimal settings by balancing exploration and exploitation.
Experimental Setup
The study utilized physical experiments to generate data for evaluation. The materials chosen were PMMA (Plexiglas hw55) and ABS (Terluran GP-35), as detailed in the table below.
Tensile bars (4mm thick) were welded and then tested in a brittle state at -40°C. This temperature was chosen to maximize the sensitivity of the longitudinal weld strength (MPa) measurements to minor variations in process quality.
Key Results and Insights
The investigation revealed a significant trade-off between peak performance and experimental cost:
- Performance vs. Effort: Bayesian Optimization achieved the highest weld strength at 50.06 MPa, but required 40 working days of experimentation. In contrast, the Potente heuristic achieved a highly competitive 49.5 MPa in just 2 working days.
- Heuristic Strength: The Potente heuristic proved to be a remarkably strong baseline, providing near-optimal results with minimal effort.
- Process Regulation: Path-regulated welding generally outperformed force-regulated welding in terms of peak strength, though it exhibited higher variance.
- Optimization Landscape: Principal Component Analysis (PCA) indicated a "rugged" optimization landscape, where small parameter changes can lead to drastic differences in strength, explaining the difficulty for simple algorithms.
Conclusion
While Bayesian Optimization provides the absolute best weld quality, its "data hunger" makes it currently impractical for many industrial applications where time-to-market is critical. The Potente heuristic remains the most efficient choice for rapid deployment. Future work should focus on "multi-fidelity" optimization—using simulations to reduce physical testing—and incorporating heuristic logic as a prior for ML models to accelerate the search for the global optimum.
Summary updated