Deep learning-assisted ultrafast ferroelectric domain imaging using atomic force microscopy

Piezoresponse force microscopy (PFM), a lock-in-based mode of atomic force microscopy (AFM), enables sensitive ferroelectric domain imaging. However, fast imaging for either massive measurements or studying fast dynamic behavior remains challenging because it suffers from a fundamental trade-off between fast acquisition and high signal quality. To address this challenge, we developed a sequential U-Net, one of the convolutional neural networks, and a transfer learning approach to reconstruct high-quality PFM images from low-quality images acquired under conditions unsuitable for reliable characterization. When applied to ferroelectric ceramics, the approach successfully recovered domain structures with more than 50% improvement compared to conventional random weight initialization and enabled PFM imaging more than 390 times faster while maintaining image quality. The framework was further validated using electrostatic force microscopy (EFM) on hafnium zirconium oxide nanocapacitors, achieving more than 67 times faster imaging, demonstrating its generalizability across lock-in amplifier-based AFM techniques. Because the proposed method requires no hardware modifications and remains compatible with standard AFM instrumentation, it offers a broadly applicable strategy for overcoming speed-sensitivity limitations in lock-in-based AFM techniques.