Lorenz Heinemann

AI-driven failure analysis in acoustic microscopy is often constrained by measurement setup dependencies, requiring models to be trained for specific configurations – such as transducer type, focus distance, or excitation energy. Furthermore, traditional workflows rely on labeled training data containing defect and non-defect scans.
This presentation introduces an anomaly detection approach that reduces the need for defect-labeled training data. By training solely on non-defective components, the model can detect deviations in new, unknown samples of the same type, enabling defect identification without prior knowledge of specific failure modes.
Beyond this, we explore methods to further generalize AI models for acoustic microscopy. Typically, even minor changes in measurement settings can significantly degrade model performance. We present initial steps to eliminate these setup-specific dependencies, allowing models to maintain accuracy across different scanning conditions – even ones, which are not included in the training data. This advancement has the potential to streamline AI-based failure analysis workflows and enhance the adaptability of machine learning models in FA.

Lorenz Heinemann is a researcher at the Fraunhofer IMWS in the department of Non-Destructive Defect Localization. He obtained his master’s degree in Computer Science from Leipzig University in 2023. His primary focus is on acoustic microscopy, where he is dedicated to advancing the field through novel AI-based analysis methods and user-friendly applications. His goal is to make acoustic microscopy more accessible and effective on software side.