Manufacturing, Ultrasonic Testing
A Numerical Simulation-Enhanced Approach
to Sorting Geometrically Complex Additively Manufactured Ceramic Parts Based on Their Resonant Ultrasonic Response
ABSTRACT
With the rise of additive manufacturing (AM), nondestructive evaluation (NDE) of complex AM parts is becoming essential. This study combines resonant ultrasound spectroscopy (RUS), finite element (FE) simulations, and machine learning to classify ceramic samples with complex geometries. Because variability among nominally identical samples renders traditional RUS analysis ineffective, a random forest (RF) model is trained predominantly on FE simulation data, supplemented by varying amounts of experimental data. Accuracy improves with more experimental training data, but exceeds 70% even when trained with less than 25% of the experimental data. The best results are achieved by selecting a narrow frequency range, guided by simulation results and RF feature importance. This method—training primarily on FE simulations and testing on experimental data—offers a practical way to interpret complex RUS data and could be applied to other data-scarce NDE scenarios.
KEYWORDS: resonant ultrasound spectroscopy, random forest, machine learning, ceramic additive manufacturing, finite element analysis
