Seismic processing is fundamental for proper subsurface imaging for resource exploration, subsurface monitoring, and Earth discovery. In recent years, supervised machine learning (ML)-based methods for seismic processing， such as denoising, have become popular approaches. However, these methods rely heavily on clean seismic labels for training, which are rarely available in field data where noise-free data is practically unattainable. Although self-supervised learning (SSL) frameworks, such as noise2void (N2V), have demonstrated success in removing random noise by learning from only the noisy inputs, they are less effective in suppressing coherent noise, such as noise with temporal dependence (trace-wise) and noise with both temporal and spatial dependence (e.g., ground roll).

 In my thesis, I first develop self-supervised denoising methods for seismic data: blind-trace mask for trace-wise noise and blind-fan mask for ground roll, two common and challenging forms of structured noise. The core idea of these blind networks lies in the design of noise masks that obscure structured noise during training, preventing the network from learning the noise itself. This guides the model to focus on reconstructing only coherent, meaningful signals. The proposed methods enable an effective self-supervised denoising framework that operates entirely on noisy seismic data, using the raw data as labels and their corrupted counterparts as inputs, to successfully suppress trace-wise noise and ground roll. Validation on synthetic and field datasets demonstrates excellent denoising performance, along with high robustness and generalizability, and only minimal signal leakage compared to conventional methods.

 While previous methods rely on manually designed masks tailored to one specific noise type, I further propose a unified and robust self-supervised framework that uses a learnable Gabor filter module to suppress both pseudo-random noise and ground roll. By embedding prior knowledge of the desired seismic signal into the Gabor filters, the network effectively distinguishes signal from noise and generalizes across diverse datasets and noise conditions, using the same noisy data as both input and target. This Gabor-based method enables significantly improved computational efficiency by requiring fewer learnable parameters and less training epochs. Tests on both synthetic and field datasets proves the method's effectiveness, showing superior denoising capability, strong robustness, and minimal signal leakage compared to traditional approaches.

 Beyond denoising, my thesis tackles another key challenge in seismic processing: salt body inversion. Salt structures are difficult to image and invert due to their high velocity contrast, complex geometries, and limited low-frequency and far-offset data. Traditional ML approaches, such as U-Net-based networks combined with full waveform inversion (FWI), often struggle to accurately reconstruct these intricate salt shapes. To overcome this, I investigate generative models as a more effective alternative. Unlike U-Nets, generative models learn the underlying data distribution from training samples, enabling them to synthesize geologically realistic salt geometries. These learned distributions are then leveraged within the salt flooding/unflooding process. This enables an efficient, fully automated workflow for challenging salt body inversion by replacing the U-Net with a generative model that leverages learned prior information about salt structure distributions to effectively predict complex salt geometries. Tests on synthetic data have demonstrated improved inversion performance under limited-offset and low-frequency conditions, underscoring the method’s strong potential for tackling complex salt body inversion challenges.

 Defense Committee: