As next-generation synchrotron facilities achieve faster, higher-resolution imaging, the amount of data produced by the experiments will also grow exponentially. Reconstructing 2-D images of the source of a diffraction pattern is already a time-consuming and computationally intensive iteration process, 3-D images provide an even bigger challenge that requires even more data and computational resources.

Chan et al. trained a neural network model to predict the 3-D shape and local strain of a crystal from its diffraction pattern in sub-second time. Their 3-D coherent diffraction imaging neural network (3D-CDI-NN) is trained on a dataset of atomistic simulations and utilizes a refinement step in its predictions, which bypasses the need for complex phase retrieval techniques that introduce computational obstacles in traditional approaches.

“The neural network model approach is a major step towards the development of near real-time data analysis tools that can rapidly screen through a large volume of collected data, and eventually evolve into techniques that can empower learning algorithms to realize autonomous diffraction imaging experiments,” said author Henry Chan.

Testing the 3D-CDI-NN on the X-ray diffraction pattern of a sample of gold nanoparticles, the authors were able to predict the shape and features of the sample in about 145 ms – hundreds of times faster than traditional iterative approaches. Notably, iterative techniques require the data to be obtained at higher rates in a process known as oversampling, but the neural network approach was able to make good predictions from undersampled data.

“More future work needs to be done to investigate this, but the ability to work with undersampled data opens new possibilities to significantly speed up the data acquisition and analysis process in experiments,” said author Mathew Cherukara.

Source: “Rapid 3D nanoscale coherent imaging via physics-aware deep learning,” by Henry Chan, Youssef Nashed, Saugat Kandel, Stephan O. Hruszkewycz, Subramanian Sankaranarayanan, Ross Harder, and Mathew J. Cherukara, Applied Physics Reviews (2021). The article can be accessed at