SAM Signal Deconvolution: Extracting Transducer independent Information
SAM Signal Deconvolution: Extracting Transducer independent Information
Tuesday, October 6, 2026: 1:10 PM
Summary:
Scanning acoustic microscopy (SAM) is essential for inspecting microelectronics, but its utility is often limited because transducer-specific hardware characteristics distort time-domain signals. This dependency makes machine learning models trained on one setup perform poorly on others. To address this, the authors propose a two-stage deconvolution framework designed to extract the true acoustic reflectivity, effectively removing hardware-induced artifacts. In the first stage, a convolutional neural network (CNN) is trained on synthetic data created by convolving point-spread functions (PSFs) with random reflectivity sequences. The second stage then refines this model using unlabelled real-world measurements through a self-supervised, reinforcement-style objective. This objective aligns signal gate metrics—such as energy and maximum amplitude—between raw data and the deconvolution output. By standardizing the input data, this approach achieves significantly sharper acoustic images and, crucially, improves the model’s ability to transfer across different SAM hardware configurations.
Scanning acoustic microscopy (SAM) is essential for inspecting microelectronics, but its utility is often limited because transducer-specific hardware characteristics distort time-domain signals. This dependency makes machine learning models trained on one setup perform poorly on others. To address this, the authors propose a two-stage deconvolution framework designed to extract the true acoustic reflectivity, effectively removing hardware-induced artifacts. In the first stage, a convolutional neural network (CNN) is trained on synthetic data created by convolving point-spread functions (PSFs) with random reflectivity sequences. The second stage then refines this model using unlabelled real-world measurements through a self-supervised, reinforcement-style objective. This objective aligns signal gate metrics—such as energy and maximum amplitude—between raw data and the deconvolution output. By standardizing the input data, this approach achieves significantly sharper acoustic images and, crucially, improves the model’s ability to transfer across different SAM hardware configurations.