How Did AI Learn to Pick Seismic Phases?

PhaseNet: phase picking as image segmentation

In 2019, Weiqiang Zhu and Greg Beroza of Stanford University introduced PhaseNet. Its most important contribution was to redefine the phase-picking problem. Earlier approaches treated picking as locating one precise arrival time within a long signal. PhaseNet instead treated it as an image-segmentation problem, classifying each sample as P-wave, S-wave, or noise. This is exactly the task U-Net was designed for. U-Net was originally developed to segment medical images, such as outlining organ boundaries in X-ray and MRI scans, and picking seismic phases is essentially the same kind of problem.

The second change is less visible but equally important. The correct arrival time of a seismic wave is a single instant, and training a model to hit that exact point is difficult because the target is sharp and sparse, so a small offset counts as almost entirely wrong. PhaseNet spreads this instant into a Gaussian curve centered on the manually labeled position, with some width on either side. This builds the uncertainty of manual labels into the training target. With a wider target, the model has a tolerance for error and trains much more easily.

The model takes a three-component waveform as input, and the U-Net outputs three probability curves for P-wave, S-wave, and noise. A rise in a curve indicates the arrival of that phase. PhaseNet was trained on decades of manually labeled data from Northern California. It is fast and accurate, and it has since become the field's benchmark: new methods are usually compared against PhaseNet first.

EQTransformer: separating three tasks

In 2020, Mousavi and colleagues from the same research group introduced EQTransformer, published in Nature Communications.

PhaseNet demonstrated that deep-learning phase picking is feasible, and EQTransformer further improved the model's ability to interpret waveforms. It introduced the attention mechanism, which was spreading rapidly through AI research at the time and later became the basis of large language models. Attention allows the model to consider both distant and nearby features when interpreting a waveform segment.

Its key design is to separate three tasks: determining whether the segment contains an earthquake, picking the P-wave, and picking the S-wave. Each task is handled by an independent output branch (head), so the tasks do not interfere with each other. The reason is that PhaseNet predicts P-wave, S-wave, and noise together as three parallel channels. A detection signal covers the whole event and is wide and long, so it carries far more weight than the narrow Gaussian peaks of P and S. When everything is learned together, the P and S signals are easily overshadowed by the detection signal. Separate branches allow each task to be learned properly. EQTransformer was trained on STEAD, a large public seismic dataset considerably larger than the one used for PhaseNet.

The problem after many models appeared

After PhaseNet and EQTransformer, many new models were published. This was a positive development, but it soon led to a practical problem: the models could not be compared fairly.

Each model used a different data format, different preprocessing, and different evaluation criteria. Determining whether model A actually outperformed model B required weeks of rebuilding both data pipelines and environments, and the data loading step alone discouraged many researchers. Without reproducible comparisons, progress in the field is difficult.

SeisBench: a common interface for models and data

In 2022, a German research team (Woollam, Münchmeyer, and colleagues) released SeisBench to address exactly this problem.

SeisBench works much like a USB interface on a computer: any trained model (such as PhaseNet or EQTransformer) and any public dataset connect in the same way, and switching models or datasets requires changing a single line of code. Environments that once took weeks to build now run with a few lines of code, and results from different researchers can be compared directly.

SeisBench did not propose a new phase picker, but its contribution is equally important. It moved the field from separate, independent development to tools that anyone can use and results that anyone can reproduce. A student new to the field can now reproduce the results of several past papers in a single afternoon, which was hardly possible five years ago.

Directions for further research

Over the past few years, AI phase picking has grown from a new idea into an open and reproducible ecosystem. The infrastructure is in place, and the barrier to entry is lower.

Mature tools, however, do not mean that every problem is solved. Models still fail in certain situations, for example when S-wave amplitudes are suppressed, or when performance drops in regions and with instruments not seen during training. With the foundation established, research has naturally shifted from feasibility toward these more specific problems.

For an analysis of one of these problems, see Why Does AI Often Miss S-Waves? →