Researchers at MIT and Argonne National Laboratory have developed a machine learning technique that can significantly accelerate the process of nanoscale X-ray imaging of integrated circuits. This could revolutionize how electronics are manufactured and tested.
Integrated circuits, or microchips, are the building blocks of modern electronics, and their continued miniaturization makes devices increasingly complex and powerful. However, as the components of these microchips shrink, it becomes more difficult to inspect and test them using conventional imaging techniques.
One promising method for imaging nanoscale components is synchrotron X-ray ptychography tomography, which uses high-energy X-rays to penetrate materials and produce detailed images of their internal structure. However, X-ray imaging is a slow process requiring precise positioning of the sample and detector, and it can take hours or even days to acquire a single reconstruction.
To speed up this process, researchers at MIT and Argonne turned to machine learning. They trained a neural network to predict the exact reconstruction of an object in a fraction of the time it would normally take. Their network, called APT or Attention Tomography, takes advantage of normalized priors in the form of patterns typical of those found inside integrated circuits and the physics of X-ray propagation through objects.
“Neural networks can learn and generalize from small amounts of data, allowing integrated circuits to be rapidly imaged and reconstructed,” said Ikson Kang, the first author of the paper. The researchers noted that their approach significantly reduced the total data acquisition and computation time required for imaging. They tested the technology on real integrated circuits and were able to capture detailed images in just minutes instead of hours.
“This new method could be an effective solution for quality assurance,” they said. “By accelerating the imaging process, we will be able to connect factories to synchrotron X-ray sources.”
The researchers noted that their approach could have profound implications for fields as diverse as materials science and biological imaging. “Our work addresses significant challenges in noninvasive X-ray imaging of nanoscale objects such as integrated circuits,” said the first author. “We believe that our physics-based attentional machine learning framework can be applied to other areas of nanoscale imaging.”
The study, titled “Attention Ptyco Tomography (APT) for Three-Dimensional Nanoscale X-ray Imaging with Minimal Data Acquisition and Computation Time,” was published in the journal Light: Science and Applications. rice field.
