Accurate subsurface velocity modeling is of fundamental scientific and practical significance for seismic data processing and interpretation. However, conventional depth-domain methods still face limitations in physical consistency and inversion accuracy. To overcome these challenges, this study proposes a deep learning-based seismic velocity modeling approach in the time domain. The method establishes an end-to-end mapping between seismic records and velocity models directly in the time domain, reducing the nonlinear complexity of mapping time-domain data to depth-domain models and improving prediction stability and accuracy. Synthetic aquifer velocity models were constructed from representative stratigraphic features, and multi-shot seismic records were generated through forward modeling. A U-Net network was employed, taking multi-shot seismic records as input and time-domain velocity fields as output, with training guided by a mean squared error (MSE) loss function. Experimental results show that the proposed strategy outperforms conventional depth-domain approaches in aquifer structure identification, velocity recovery, and interlayer contrast depiction. Quantitatively, significant improvements in MSE, peak signal-to-noise ratio, and structural similarity index indicate higher reconstruction reliability. Overall, the results confirm the effectiveness and potential of the proposed time-domain framework for aquifer velocity inversion and its promise for intelligent seismic velocity modeling.
9 February 2026 in IEEE Transactions on Geoscience and Remote Sensing
Deep unfolding networks (DUNs), which offer strong interpretability, have gradually attracted increasing attention in seismic data reconstruction. Most existing DUN approaches are limited to unfolding in the time-space (t–x) domain, making it difficult to fully exploit the multiscale features of seismic signals, thereby decreasing reconstruction performance and often leading to the loss of fine details. To overcome the above limitation, a dual-domain optimization framework is proposed, in which the traditional iterative process is replaced by neural networks, yielding a dual-domain deep unfolding network (D3U-Net) based on implicit regularization. The proposed method combines prior information from both the t–x domain and the wavelet domain, guiding the network to simultaneously capture spatial structural information and frequency characteristics at each stage, thus achieving cooperative reconstruction in both domains and enhancing detail recovery. Meanwhile, an implicit regularization mechanism is introduced, where a learnable neural network substitutes the conventional proximal mapping operator and does not require manual specification of the weights of the regularization term, simplifying the tuning process and endowing the network with greater flexibility and adaptability. Experiments conducted on synthetic data, the BP2004 dataset, and field marine seismic data demonstrate that D3U-Net achieves superior reconstruction accuracy and stability under both regular and random seismic-trace missing scenarios compared with current state-of-the-art methods, confirming overall advantages in reconstruction fidelity, generalization capability, and interpretability.