Why Does AI Often Miss S-Waves?

For this reason, seismologists have in recent years assigned this work to deep learning. After training, a model takes a waveform segment as input and outputs the arrival times of the P-wave and the S-wave.

P-waves and S-waves are two types of seismic waves. P-waves travel faster and arrive first. S-waves travel more slowly and arrive later, but they are more destructive and cause most of the noticeable shaking in buildings. Both arrival times are essential for determining the location and magnitude of an earthquake.

An unexplained failure mode

The problem is that AI performs consistently when picking P-waves but often fails on S-waves. On the same waveform, where an analyst can clearly see the S-wave, the model outputs a distorted peak of insufficient height that falls just below the detection threshold, so the S-wave is counted as undetected.

Notably, the position the model identifies is usually correct. The timing is accurate, and only the height is insufficient.

Seismologists have been aware of this phenomenon for years. Common responses include using larger models, adding more data, or adjusting the training method, treating the problem as an inconvenient side effect to be avoided.

This paper does not avoid the problem. Instead, it sets out to identify the underlying cause.

Why predictions stall

Consider judging how closely a portrait resembles its subject. The intuitive approach is to look at the whole face. If the rules instead prohibit viewing the whole image and require dividing it into hundreds of small cells, each scored independently and then summed, the judgment becomes difficult.

Each cell can only indicate whether it should be brighter or darker, and no cell can indicate that the whole face should move to the left. When the predicted shape is correct but its position is slightly to the right, each cell receives only one of two signals: brighten here or lower here. The prediction therefore becomes fainter, while its position is never corrected.

The conventional way of training AI to pick S-waves is exactly this kind of cell-by-cell scoring. It has no mechanism to tell the model which direction to move as a whole and can only adjust values up or down. As a result, the S-wave prediction stalls at one position and becomes progressively flatter.

S-waves are particularly susceptible to this problem because of two additional factors. First, the ground truth for S-waves is inherently less precise. As S-waves travel farther, the waveform broadens and its onset becomes less distinct, so even expert picks can differ by several hundred milliseconds, and the correct answers in the training data contain noise. Second, AI models tend to respond to sharp features and anchor the prediction to the most prominent change in the waveform, which is not where the S-wave actually is.

Together, these three factors leave the S-wave prediction stalled in a state that training cannot improve.

The solution

Once the cause was identified, a solution could be designed accordingly.

The authors used a technique called a GAN. GANs are usually used to generate images or videos. Two models are trained against each other, one generating outputs and the other discriminating between real and generated data. The generator tries to fool the discriminator, the discriminator tries to distinguish real from generated, and the generator's outputs become increasingly close to real data.

In this paper, the discriminator examines the model's output and judges whether its overall shape resembles a real label. It evaluates the overall shape of the output rather than scoring each cell individually.

This design supplies the signal that was previously missing. The model must not only get each cell's value right but also produce a plausible overall shape, or the discriminator will detect the difference.

Significance beyond the 64%

The 64% gain is a striking figure, but the more notable aspect of this paper is its method.

The paper takes a phenomenon that had been treated as an engineering issue and breaks it down into something that can be understood, predicted, and analyzed. It explains why predictions stall, where they stall, and how to release them, which turns the process from trial and error into principled analysis.

This diagnostic approach is not limited to seismology. It applies to any task that marks points in time within a signal, such as heartbeats in an ECG, events in satellite signals, and syllable boundaries in speech. Whenever the ground truth is a signal with a shape, conventional point-wise scoring can lead to the same problem.

The paper concludes with one point: understanding the shape of the loss function is as important as understanding the model architecture. Researchers often discuss which network to adopt, how many layers to add, or which function to use, but rarely examine whether the objective the model optimizes is itself well designed. If the scoring rule is flawed, a more powerful model cannot solve this type of problem.