Abstract: This paper develops an efficient calculation model for the dam seepage flow based on an improved conditional deep convolution generative adversarial network to overcome the problem of previous surrogate models in numerical simulation of the flow field. A traditional surrogate model is mostly constructed based on local monitoring points, but it is time-consuming, computationally intensive, and difficult to capture the overall seepage flow features at key cross-sections, thereby failing to meet the needs of rapid engineering visualization and decision-making. This new model achieves efficient prediction through constructing a mapping relationship between working conditions and seepage flows at key cross-sections. We apply a squeeze-and-excitation (SE) channel attention mechanism and residual networks to the generator, so as to improve its feature extraction, and integrate the discriminator with the Haar wavelet transform to strengthen its edge information recognition and improve its distribution feature capturing. In addition, super-resolution techniques are incorporated to reconstruct high-resolution seepage fields. Case studies demonstrate our new model achieves significant improvement on efficiency over traditional numerical methods. Compared with the unmodified Generative Adversarial Network (GAN), it achieves an average increase of 44.83% in Fréchet distance, 2.54% in structural similarity index, and 4.25% in peak signal-to-noise ratio, validating its effectiveness and superiority.

Key words: dam seepage field, conditional deep convolutional generative adversarial network, channel attention mechanism, Haar wavelet transform, super-resolution, residual network