Abstract: The need for dense spatial sampling can be alleviated by predicting room impulse responses (RIRs) at additional locations in the space from measurements obtained using limited microphones. In this paper, a data-driven approach for RIR reconstruction is proposed. The proposed method utilizes an improved time domain U-Net network, which enhances the ability of the network for feature learning through the optimization of the fundamental blocks. In the proposed method, the control region is divided into square blocks. The RIRs at the four vertices of each square are inputted into the proposed U-Net network, which aims to predict the RIR at the central position of the square. The converged U-Net network can reconstruct RIRs for supplementary positions. Experiments conducted on a recorded RIR dataset containing multiple room scenarios demonstrate that the proposed method outperforms the traditional singular value decomposition technique in terms of the RIR reconstruction performance. Additionally, the performance of the proposed U-Net network surpasses that of the original U-Net, CRN and DCCRN networks. In a practical scenario concerning sound zone reproduction, the utilization of the proposed method for predicting RIRs has resulted in higher acoustic contrast between the listening and quiet zones with fewer microphones. This further validates the merits of the proposed approach.