Physics-informed unsupervised learning for sparse-view tomography imaging in plasma diagnostics

Abstract

In high-energy-density physics, plasma density distribution dictates laser coupling efficiency and drive symmetry. However, Real-time 3D diagnostics of plasmas remain challenging, primarily constrained by their large geometries and ultrafast dynamic evolution. Here, we propose a physics-informed deep learning method for common-path interferometric tomography imaging technique, which resolves this challenge by integrating physical modeling with an unsupervised learning strategy. Technically, a beam splitter cube is utilized to construct a lens-free common-path interferometric configuration, enabling high-precision phase recovery. By employing a pair of oppositely placed trapezoidal prisms to converge three parallel beams onto the sample and subsequently re-collimate them into parallel paths, where each beam interferes via a beam splitter cube, the rapid, simultaneous capture of three-angle object information is achieved through a single-shot measurement. To tackle the severe information deficiency caused by sparse-angle projections which are insufficient for traditional reconstruction, an unsupervised deep learning algorithm is employed to solve for the optimal solution, thereby achieving accurate 3D tomographic imaging reconstruction. Results from both simulations and experimental demonstrate that this method facilitates rapid and effective real-time 3D tomographic reconstruction from extremely sparse angles (only three angles) via single-shot measurement. The successful development of this technology will provide valid diagnostic support for enhancing the performance of intense radiation sources and validating sophisticated physical models.