Enhancement of Scanning Acoustic Microscopy in Robustness, Precision and Automation through AI-based Signal Analysis

Abstract Inspection techniques of non-destructive character are highly relevant in a variety of fields in scientific and industrial applications. Due to the avoidance of sample alteration and destruction they allow for fault isolation in even highly complex microelectronic devices for failure analysis, technology development and quality assessment. The present paper focuses on the extension of scanning acoustic microscopy, which although highly sensitive towards structural variation like cracks, fractures and delamination, requires extensive interpretation of the acquired echo signals for deriving conclusive and correct interpretations with sufficient certainty. Previous research has proven the value and potential of supplementing acoustic signal analysis by approaches of machine learning. However, while proving them highly effective the data preparation and handling of the investigated approaches under practical conditions concerning robust and reliable deployment was not a priority. This previous research was focused on investigating general applicability, technical benefit, effectiveness as well as the limitations in performance. The main focus of the present paper is on an extended two-stage ML approach for simplifying and automating the identification and selection of the regions in the acquired acoustic echo signals that carry the relevant information for assessing the inspected parts' conditions. The developments described here avoid the sensitivity of the classification and detection performance on the temporal position of the signal features in the underlying time-domain signals. With the described concept, automation, robustness, precision and reliability of the acoustic signal analysis and the subsequent defect identification can largely be increased.