New Ptychography Networks Improve Out-of-Distribution Generalization

A new research paper introduces an advancement in ptychography neural networks, addressing a significant limitation: their poor generalization to out-of-distribution data due to scaling inconsistencies. The proposed solution involves a novel factorization strategy that effectively separates the learned object texture from the scaling of the measurement data. This allows a single trained network to produce accurate and consistent reconstructions even when illumination conditions vary significantly. The key to this improvement lies in predicting the learned object in real and imaginary units, rather than the conventional amplitude and phase representation. Additionally, the researchers developed a synthetic object sampling strategy specifically designed to minimize phase distribution mismatch between the synthetic training data and real experimental targets. These innovations collectively lead to a substantial reduction in Fourier error, up to five times, compared to previous baselines across multiple experimental datasets from various beamlines and facilities.