AI Enables Real-Time Coherent Imaging

Ptychography creates nanoscale images of macroscopic objects but generates data faster than many methods can process it. This AI workflow processes data at the edge in real-time at speeds up to 2,000 frames per second while using less radiation.

Ptychography produces high-resolution images by scanning a coherent beam across a sample and recording diffraction patterns at each position. The technique achieves nanoscale resolution for centimeter-sized objects without destroying samples, making it valuable for studying integrated circuits, biological specimens, and structural materials. However, modern detectors capture data faster than traditional computational methods can reconstruct images. A single scan across one square millimeter generates 200 terabytes of data at 16 gigabits per second, requiring petaflops of computing power for conventional analysis. This data bottleneck prevents scientists from seeing results during experiments or adjusting measurements based on what they observe.

Researchers at Argonne National Laboratory developed a workflow that achieves real-time imaging by combining artificial intelligence at the beamline with high-performance computing resources. The system uses supercomputers to train neural networks on incoming data, then deploys updated models to a palm-sized edge device that processes individual diffraction patterns as the detector captures them. During demonstrations at the Advanced Photon Source, the workflow processed data streaming at 2,000 frames per second while running on low-power hardware at the beamline. The neural network analyzes each diffraction pattern independently in milliseconds, but also reducing the amount of data need to be collected.

This approach enables low-dose imaging by removing oversampling constraints. Traditional ptychography requires 90% overlap between measurement points to reconstruct images accurately. The AI workflow achieved greater than 90% accuracy with zero overlap between points, reducing radiation dose by a factor of six. By processing data with shorter exposure times, the system achieved further dose reductions up to 10,000-fold compared to conventional methods. These capabilities make the technique particularly valuable for radiation-sensitive samples including biological materials and certain electronic components.