Ning Wang, Zhiyuan Lin, Bingchuan Geng, Ying Shi
ABSTRACT Full‐waveform inversion (FWI) is a powerful tool for high‐resolution subsurface imaging but remains challenged by high computational demands and strong dependence on the initial model. We propose a ConvNeXt‐based reparameterized FWI framework that revisits convolutional network design for physics‐driven inversion. The framework leverages hierarchical convolutional representations to improve multiscale feature learning while maintaining favourable computational efficiency. Comparative experiments on standard benchmark models demonstrate that the proposed method achieves inversion accuracy comparable to U‐Net and Transformer‐based approaches, with more stable convergence and reduced computational cost. The method further exhibits enhanced robustness in the presence of strong noise and limited low‐frequency data, highlighting its ability to extract consistent features across frequencies. These results indicate that modernized convolutional architectures provide an effective and practical solution for high‐fidelity seismic inversion.