Stability Modeling for Chatter Avoidance in Aerospace Machining: An Application of Physics-Guided Machine Learning

The fusion of physics-based and data-driven models represents a transformative breakthrough in data science for machining broadly, and for chatter avoidance in aerospace machining in particular. Physics-guided machine learning (PGML) leverages experimental data generated during the machining process while incorporating decades of theoretical process modeling efforts and domain knowledge. This new approach to stability modeling helps assure simultaneous achievement of part accuracy, high surface finish and productivity. The true stability model is typically unknown in practice, making optimal parameter selection a challenge. Further, machining parameters that yield stable dynamics depend not only on the time-varying dynamics of the specific machine-tool-workpiece, but also on specific conditions in the ambient operational environment. From a practical perspective, operators tend to adjust spindle speeds downwards when encountering chatter which reduces productivity. Knowledge of the true, as opposed to physics-based or manufacturer-provided, stability model would enable operators to better select and adjust machining parameters. From a theoretical perspective, PGML addresses limitations of purely physics-based or machine learning stability models. Data-driven machine learning models are typically black box models that do not provide deep insight into the underlying physics and can yielding solutions that violate physical laws or operational constraints. In addition, acquiring the large amounts of data needed for machine learning can be costly. On the other hand, many physical processes are not completely understood by domain experts so physics-based models must make simplifying assumptions that compromise optimal parameter selection. In this research we demonstrate that the accuracy of a PGML model trained using data generated from an uncertain physics-based model, and subsequently updated with experimental data and domain knowledge, exhibits convergence towards the underlying true stability model and reduces uncertainty in parameter section with reduced investment in costly data collection.