PtychoX: differentiable ptychographic operators powered by CUDA

Ptychography and related computational imaging techniques are widely used for quantitative phase reconstruction in optical and photonic systems. However, the development of advanced reconstruction algorithms is often limited by the lack of high-performance differentiable software frameworks that integrate physical forward models with modern optimization and learning-based workflows. Here we present PtychoX, an open-source, CUDA-accelerated software library for differentiable ptychographic imaging. PtychoX implements multiple ptychographic modalities through modular forward and adjoint operators written directly in CUDA and exposed to MATLAB via custom differentiable-function interfaces. By Combining GPU-optimized operations, PtychoX enables efficient end-to-end gradient computation for large-scale ptychographic reconstruction and joint optimization problems. Systematic benchmarks across conventional ptychography, coded ptychography, near-field ptychography, Fourier ptychography, and 3D multi-slice reconstruction demonstrate up to 50× acceleration over conventional MATLAB gpuArray-based implementations while maintaining reconstruction fidelity and numerical consistency. By providing a unified, differentiable, and GPU-accelerated framework for ptychographic imaging, PtychoX lowers the barrier to developing and testing advanced reconstruction algorithms and establishes a practical foundation for future research at the intersection of coherent imaging, inverse problems, and learning-based optical computation.