A Breakthrough in Seismic First-Break Picking: AI Meets Geophysics for Unprecedented Accuracy
In the relentless pursuit of energy resources beneath the Earth’s complex surface, the oil and gas industry faces a persistent, data-intensive challenge: accurately identifying the very first whisper of a seismic wave as it returns from the subsurface. This critical step, known as first-break picking, is the cornerstone of near-surface static correction, a process essential for transforming raw seismic recordings into clear, interpretable images of underground geological structures. For decades, this task has been a bottleneck, relying heavily on skilled human interpreters to meticulously scan vast datasets—a process that is not only time-consuming and expensive but increasingly unsustainable as seismic surveys grow larger and denser. The era of big data in geophysics has arrived, demanding a paradigm shift. A groundbreaking study, published in the prestigious journal Shiyou Diqiu Wuli Kantan (Petroleum Geophysics Exploration), offers precisely that shift: a novel, highly efficient method that marries the pattern-recognition power of deep learning with the physical constraints of seismic wave propagation, setting a new standard for automated first-break picking.
The research, spearheaded by David Cova and Yang Liu from the State Key Laboratory of Petroleum Resources and Prospecting at China University of Petroleum, Beijing, in collaboration with colleagues from the CNPC Geophysical Key Laboratory and BGP Inc., addresses the core limitations of both traditional and early AI-based approaches. Traditional methods, such as those based on energy ratios or signal correlation, are notoriously sensitive to noise and require high signal-to-noise ratios and consistent waveforms—conditions rarely met in the rugged, geologically complex terrains where accurate statics are most needed. Early deep learning models, while faster, often sacrificed precision, particularly in noisy environments, or were too rigid to handle the variable sizes of seismic datasets. The team’s innovation lies not in discarding these technologies, but in their intelligent synthesis and enhancement, creating a robust, end-to-end solution that is as scientifically rigorous as it is computationally efficient.
At the heart of this new methodology is the U-Net, a sophisticated type of convolutional neural network originally designed for biomedical image segmentation. The U-Net’s architecture, featuring a contracting path to capture context and a symmetric expanding path for precise localization, is uniquely suited to the task of identifying the subtle boundary between seismic noise and the first arrival of coherent energy. The researchers didn’t stop at simply adopting U-Net; they embarked on a meticulous journey of optimization and validation. They rigorously compared the standard U-Net against three of its more complex variants: UNet++, Wide U-Net, and Attention U-Net. The results were counterintuitive and profoundly instructive. Contrary to the common assumption that more complex models yield better results, the standard U-Net, when properly tuned, outperformed its fancier siblings. UNet++, with its nested, dense skip connections, offered improved feature fusion but at the cost of a 35% longer training time and no significant gain in accuracy. Attention U-Net, designed to focus on salient features, was faster to train due to fewer parameters but suffered from poor continuity in the picked first breaks, particularly at varying shot-receiver distances. Wide U-Net, which simply enlarged the convolutional kernels, became overly complex, degrading the continuity of the picked events. This finding underscores a crucial principle in applied AI: elegance and optimization often trump brute-force complexity. The key to success was not architectural extravagance, but the careful selection of hyperparameters.
The team’s hyperparameter optimization process was exhaustive and scientifically meticulous. They systematically tested and tuned every critical lever in the neural network’s learning process. For the activation function, which introduces non-linearity into the model, they moved beyond the standard ReLU, which suffers from the “dying ReLU” problem where neurons can become inactive. They evaluated Leaky ReLU, ELU, and SELU, ultimately selecting SELU for its superior performance. SELU’s ability to self-normalize its outputs to zero mean and unit variance led to faster convergence and significantly fewer prediction anomalies, making it ideal for the non-stationary nature of seismic data. For the loss function, which measures the discrepancy between the network’s predictions and the ground truth, they confronted the fundamental issue of data imbalance. In a typical seismic shot gather, pixels representing noise or coherent energy vastly outnumber those representing the thin, critical line of the first break. Standard loss functions like Binary Cross-Entropy (BCE) treat all errors equally, which is suboptimal for this highly imbalanced task. After testing a suite of advanced functions—including Dice, Tversky, and Focal Tversky, which are designed for region-based segmentation—they found that the Huber loss function delivered the best compromise. Huber loss, which is less sensitive to outliers than mean squared error, minimized prediction anomalies while only slightly sacrificing the continuity of the picked first breaks, making it perfectly suited to the noisy reality of field data.
The choice of optimizer, which dictates how the network updates its internal weights during training, was equally critical. They tested five leading optimizers: SGD, RMSprop, Adadelta, Adam, and Adamax. SGD, the most basic, performed poorly, prone to getting stuck in local minima. RMSprop and Adadelta showed improvement, but it was Adam and its variant, Adamax, that truly shined. Adam, which combines the benefits of momentum and adaptive learning rates, nearly eliminated prediction anomalies. Adamax, based on the infinity norm, went a step further, not only eliminating anomalies but also ensuring perfect continuity in the picked first breaks. For regularization, a technique to prevent overfitting, they found that a modest dropout rate of 0.05—randomly deactivating just 5% of neurons during training—was optimal. Higher rates harmed the network’s ability to learn the subtle features of the first break. Finally, for weight initialization, which sets the starting point for the network’s learning, they determined that Glorot normal initialization, which balances the variance of activations across layers, provided the most stable and effective training dynamics for their specific U-Net and SELU configuration. This granular, scientific approach to hyperparameter tuning transformed a good model into an exceptional one.
However, even the most finely tuned neural network is only as good as the data it is fed. Recognizing this, the researchers developed a novel, multi-stage amplitude balancing workflow specifically designed to prepare seismic data for deep learning. Raw seismic data is plagued by issues like spherical divergence, where signal energy decays with distance, and erratic noise bursts, which can confuse a learning algorithm. Common industry practices like Automatic Gain Control (AGC) often exacerbate the problem by amplifying pre-first-break noise, making it harder for the network to distinguish signal from noise. Root Mean Square (RMS) normalization is better but can still leave deep, weak signals under-enhanced and distant traces over-amplified. The team’s solution is a carefully orchestrated sequence: first, apply T-squared compensation to counteract spherical divergence; then, clip amplitudes to the 99th percentile to tame extreme outliers; next, apply clipping based on the Interquartile Range (IQR), a robust statistical measure less sensitive to outliers than standard deviation, to further stabilize the data; followed by trace-by-trace RMS normalization to balance energy across individual seismic traces; another round of IQR clipping to handle any new outliers introduced by normalization; and finally, min-max scaling to confine all amplitudes to a 0-1 range, which is ideal for the neural network’s internal operations. This proprietary workflow doesn’t just make the data look better; it fundamentally reconditions it, creating a balanced, noise-resilient input that allows the U-Net to learn the true characteristics of the first break with remarkable clarity.
The final, and perhaps most ingenious, component of their method is the integration of geophysical constraints. A neural network, no matter how well-trained, operates in a statistical vacuum. It doesn’t inherently understand that seismic waves must obey the laws of physics. To bridge this gap, the team introduced a “velocity constraint” step applied after the U-Net generates its initial segmentation map. In near-surface geology, seismic velocities typically fall within a known, reasonable range—for their study area, they defined this as 1650 to 2100 meters per second. The algorithm scans the segmentation map and identifies pixels that, based on their position relative to the source, would imply a physically impossible velocity (either too fast or too slow). These “rogue” pixels, which are almost always misclassifications caused by noise, are then corrected. Pixels above the first break that suggest impossibly high velocities (likely noise misclassified as signal) are set to “noise.” Pixels below the first break that suggest impossibly low velocities (likely signal misclassified as noise) are set to “signal.” This simple, physics-based post-processing step acts as a powerful error-correction mechanism. The results were quantifiable: applying the velocity constraint boosted the Dice coefficient—a key metric for segmentation accuracy—from 0.9854 to 0.9862 and the Jaccard index from 0.9723 to 0.9731. More importantly, it dramatically improved results in the most challenging areas: the far-offset, low signal-to-noise regions where traditional methods often fail. Here, the average picking error was reduced from 8.7 milliseconds to just 5.2 milliseconds, a significant improvement for seismic processing.
The practical validation of this method is compelling. Tested on a large land-based 2D seismic dataset comprising 1,000 shot gathers, the optimized U-Net pipeline delivered exceptional results. It achieved a staggering training accuracy of 99.83% and a validation accuracy of 99.79%, demonstrating its ability to learn and generalize. When compared head-to-head against industry-standard methods like STA/LTA and energy ratio, as well as a traditional CNN, the superiority of this new approach was undeniable. While the STA/LTA method was the fastest, its average picking error was 46.5% higher. The energy ratio method was slow and inaccurate. The traditional CNN performed well but still had a 31.1% higher error rate than the new method. The proposed method achieved an overall average absolute error of just 4.08 milliseconds, with processing times of only 1.2 seconds per shot gather—a speed that makes it viable for processing modern, high-density surveys with millions of traces. Visually, the picked first breaks were smooth, continuous, and aligned perfectly with manual picks in high signal-to-noise areas, while maintaining remarkable accuracy even through zones of ground roll and in areas with significant topographic relief.
This research represents more than just an incremental improvement; it is a blueprint for the future of seismic data processing. It demonstrates that the most effective AI solutions are not black boxes, but carefully engineered systems that respect and incorporate domain-specific knowledge. By combining a rigorously optimized deep learning model with intelligent data pre-conditioning and physics-based post-processing, the team has created a method that is accurate, robust, efficient, and, crucially, interpretable. It sets a new benchmark for automated first-break picking and paves the way for similar hybrid approaches in other areas of seismic interpretation, from horizon tracking to fault detection. The implications for the industry are profound: faster project turnaround times, reduced operational costs, and most importantly, higher-quality subsurface images that lead to better drilling decisions and more successful exploration outcomes.
The study, titled “Artificial Intelligence and Apparent Velocity Constrained Seismic First-Break Picking Method,” was authored by David Cova, Yang Liu, Chengzhen Ding, Chenglin Wei, Fei Hu, and Yunzhu Li. The authors are affiliated with the State Key Laboratory of Petroleum Resources and Prospecting, China University of Petroleum, Beijing; the CNPC Geophysical Key Laboratory, China University of Petroleum, Beijing; the School of Petroleum, China University of Petroleum at Karamay; and the Research Institute of BGP Inc., Zhuozhou, Hebei. It was published in Shiyou Diqiu Wuli Kantan (Petroleum Geophysics Exploration), Volume 56, Issue 3, pages 419-435, in 2021. The DOI for this landmark publication is 10.13810/j.cnki.issn.1000-7210.2021.03.001.