Machine learning-driven ultrasonic monitoring for quality assurance in additive manufacturing employing 2D phononic coupons

Abstract

Additive Manufacturing/3D Printing (AM/3DP) is a critical fabrication technology, especially for manufacturing high-value, high-performance, complex parts and components requiring unprecedented degrees of morphological (shape) complexities, material combinations, and internal geometrical intricacies. However, in AM/3DP, quality assurance remains a persistent problem. AM/3DP processes are known to be volatile and involve coupled fields with strong nonlinearities; as a result, traditional first-principles-based approaches often struggle with accurate process modeling and predictions. Ultrasound is a promising Non-destructive Evaluation (NDE) technique, but its direct use has limitations in evaluating compact parts with intricate internal structures due to the complex diffraction and reflection fields that its interfaces cause in the near-field, leading to challenges in building principles-based mathematical models. The current proof-of-concept study introduces a Machine Learning (ML)-driven ultrasonic monitoring framework using Phononic Test Coupons (PTCs) for real-time quality assurance in AM/3DP. PTCs are specially designed lattice-based coupons with periodic internal structures that mimic the actual build's critical geometric and structural attributes. In the presented model PTC designs, each layer consists of 20 closely packed parallel lines (bundles) printed in a Fused Filament Fabrication (FFF) process from PLA material, stacked in a six-layer structure totaling 120 bundles. Each bundle is labeled with a unique global bundle index. A Deep Neural Network (DNN) model that detects and localizes defects in real-time by analyzing the experimental ultrasonic elastic waves acquired during printing is developed and used. The effectiveness of the resulting DNN model is tested with previously unseen experimental data, demonstrating an accuracy of over 86 % in predicting the global bundle index from the ultrasonic waveform. The proposed framework could provide real-time feedback to the closed-loop control system of the machine or proactively halt defective prints, allowing for necessary adjustments in the process.