研究目的
To adapt automatic differentiation framework to solve practical and complex ptychographic phase retrieval problems and demonstrate its advantages in terms of speed, accuracy, adaptability and generalizability across different scanning techniques.
研究成果
The paper concludes that Automatic Differentiation Ptychography (ADP) offers significant advantages over current state-of-the-art algorithms like PIE (ePIE and multi-probe ePIE) in terms of speed and accuracy. ADP's flexibility in adapting to modifications in the forward model and its superior convergence properties are demonstrated through both synthetic and real experimental data. The study also highlights the importance of optimizing minibatch sizes for balancing speed and accuracy in ADP.
研究不足
The study acknowledges the practical difficulty in determining the exact complex beam profile of the probe in ptychography experiments, which affects the accuracy of probe and object reconstruction. It also notes the trade-off between speed and accuracy with minibatch size in ADP.
1：Experimental Design and Method Selection:
The study employs automatic differentiation (AD) to solve ptychographic phase retrieval problems, comparing its performance with traditional methods like PIE and ePIE. The methodology involves formulating the phase retrieval problem as an optimization problem and using AD to automatically compute gradients for optimization.
2：Sample Selection and Data Sources:
Simulated and real experimental data from the Bionanoprobe at the 21-ID-D beamline for the Advanced Photon Source (APS) are used. Simulated data includes integrated-chip images for step-scan and fly-scan modes.
3：List of Experimental Equipment and Materials:
Nvidia TITAN X GPU for computational speed comparison, simulated chip images of size 512×512 for step-scan and 256×256 for fly-scan, with specific beam diameters and probe sizes.
4：Experimental Procedures and Operational Workflow:
The study involves running both ADP and ePIE algorithms on the same computational resource, optimizing parameters like minibatch size and learning rate for ADP, and comparing reconstruction quality using PSNR metrics.
5：Data Analysis Methods:
The quality of reconstruction is compared using Peak Signal to Noise Ratio (PSNR) between the reconstruction and ground truth images. The computational speed is compared by running both algorithms on a Nvidia TITAN X GPU for fixed time.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
