A Review of Deep-learning-based Seismic Data Denoising and Its Promising Paradigm Shift to Foundation Models

Beneath the Earth's surface lies a complex history of rock layers, faults, and hidden resources, but listening to that history is often like trying to hear a whisper in a hurricane. Geophysicists use sound waves to map these underground structures, sending vibrations down and recording how they bounce back. However, the journey of these waves is rarely clean. As they travel, they are battered by a chaotic mix of interference: the rumble of distant traffic, the vibration of wind, the erratic spikes of electrical equipment, and the natural, rolling energy of the ground itself. This background noise drowns out the faint, valuable signals that reveal the true shape of the subsurface. For decades, scientists have tried to filter out this noise using mathematical rules designed for specific types of interference, but the Earth is too varied for a single rule to work everywhere. When the noise changes, the old filters often fail, sometimes stripping away the very signals they were meant to protect.

A team of researchers from Jilin University and Northeast Electric Power University has proposed a different way to solve this problem, moving away from building a new tool for every specific noise type and toward creating a single, highly adaptable system. They developed a new approach called a "foundation model," a concept borrowed from how computers learn to understand language or recognize images. Instead of teaching a computer to recognize only one kind of noise, they taught it the fundamental language of clean seismic signals. By exposing the system to massive amounts of clean data first, the model learned the intrinsic patterns of how seismic waves naturally behave. Once it understood the "voice" of the Earth, the researchers showed that it could quickly learn to ignore specific types of noise with very little additional instruction. Their work, centered on a system they named SeisDeFM, demonstrates that this method can strip away erratic, external, and random noise while preserving the delicate, deep reflections that traditional methods often damage.

The core of this new method lies in a two-step learning process that mirrors how a human might learn a skill. First, the system undergoes a massive pre-training phase. Imagine a student spending years reading every book in a library to understand the structure of language, the flow of sentences, and the logic of stories, without ever being asked to write a specific essay. In this stage, the SeisDeFM model studied thousands of clean seismic records, learning to recognize the continuous, coherent patterns of real geological events. It learned to distinguish the smooth, logical flow of a reflection from the chaotic, broken nature of noise. This phase was self-supervised, meaning the model taught itself by trying to reconstruct incomplete data, forcing it to understand the deep connections between different parts of a seismic record.

Once the model had mastered the general language of seismic waves, it moved to the second stage: adaptation. This is where the system learns to handle specific problems. In the past, if a geophysicist wanted to remove a specific type of noise, like the rolling waves from the surface, they had to train a new model from scratch using thousands of examples of that exact noise. This was slow and required vast amounts of labeled data. With SeisDeFM, the researchers only needed a tiny fraction of that data. They took the pre-trained model and showed it just a few thousand examples of noisy data paired with the clean version. Because the model already understood the underlying structure of the signals, it could rapidly learn to identify and subtract the new noise. In their tests, the model achieved superior results using only one-tenth of the labeled data required by traditional methods, proving that a deep understanding of the signal allows for much faster and more efficient learning of the noise.

The researchers tested this system against four distinct and difficult types of noise found in real-world field data. The first was erratic noise, which appears as sudden, high-amplitude spikes caused by things like recording glitches or wind. Traditional methods often struggle here, either leaving the spikes behind or accidentally removing the weak signals that happen to be near them. The new model, however, successfully removed these spikes while keeping the faint, deep reflections intact. The second test involved external noise from human activity, such as vehicle movement, which can smear the data. Again, the foundation model cleaned the data more thoroughly than previous methods, preserving the continuity of the geological layers. The third challenge was surface waves, a powerful, low-frequency noise that often masks the very first reflections from the ground. The model suppressed these waves completely without blurring the sharp details of the underlying rock. Finally, they tested it against random noise, the static-like interference that covers a broad range of frequencies. The model removed this noise while maintaining the spectral integrity of the signal, ensuring that the frequency content of the Earth's response remained accurate.

A key finding of the study is that the model's success depends on combining two different ways of looking at the data. The researchers built the system with a dual architecture: one part that looks closely at small, local details, and another that steps back to see the big picture across the entire record. The local part is excellent at catching the specific shape of a noise spike, while the global part understands how a seismic event stretches across time and space. By merging these perspectives, the model avoids the common mistake of confusing a weak, continuous signal with noise. The study also showed that while the model requires a significant amount of computing power to learn the general language of seismic waves initially, the actual process of applying it to a new job is incredibly fast and efficient. It does not need to be retrained from scratch for every new location or noise type; it simply adapts its existing knowledge.

The results suggest a shift in how seismic data might be processed in the future. Instead of building a new, specialized tool for every new noise problem, the industry could rely on a single, robust foundation model that has learned the fundamental physics of the Earth. This approach not only saves time and computational resources but also improves the quality of the final image by reducing the risk of accidentally deleting valuable data. The researchers demonstrated that by teaching a machine to understand the "truth" of a clean signal first, it becomes far better at recognizing and removing the "lies" of noise. This method offers a promising path forward for handling the massive volumes of data collected in modern exploration, turning a chaotic, noisy recording into a clear, high-fidelity map of the subsurface.