Deep-Learning-Based High-Efficient Sparse Array Design for a K-Band Short-Range MIMO Radar System

Sparse virtual array design is essential for improving the resolution of multiple input multiple output (MIMO) radar systems by enlarging the array aperture. However, its design typically relies on ambiguity function (AF) analysis, which, although accurately characterizing spatial beamforming properties, is hindered by the heavy computational cost of manifold matrix operations, leading to long optimization time. To address this challenge, a deep-learning-based ambiguity function generator (AFG) is introduced to directly predict AFs from virtual-array configurations and directions of arrival, eliminating explicit AF computation and enabling efficient sparse MIMO array design. Based on AFG, a sparse array was implemented on a custom 2T4R K-band MIMO radar, achieving a 32.8% improvement in spatial resolution while maintaining a −10 dB sidelobe level across a 100° field of view. The AFG delivers over 35× computational speedup with AF prediction RMSE below 0.1, establishing a fast, accurate, and cost-effective paradigm for MIMO radar array design.