San Andreas Fault: Hidden movements revealed by artificial intelligence

Pipeline architecture

Input data consists of continuous daily borehole strain observations and their wavelet representations derived from corrected strain components. The neural network is an autoencoder with skip connections that learns a compact latent representation of the input. A downstream unsupervised clustering step then partitions the latent space to separate deformation signals from non-deformation artifacts. The skip connections are the load-bearing design choice: they preserve high-frequency transient structure across the bottleneck, consistent with the target signal class — short-duration slow slip events that a standard encoder-decoder would attenuate during compression. The figure in the source material shows white input curves, wavelet coefficients as background colors, and highlights a representative event recorded on 15 June 2013.

Empirical result

The reported finding is a previously hidden population of short-duration slow slip events at Parkfield, systematically followed by elevated low-frequency earthquake activity. The temporal coupling between silent slip and low-frequency seismicity is the substantive scientific claim; the neural network is the detection substrate that made the population enumerable. Parkfield is one of the most intensively monitored fault segments globally, and prior instrumentation had not surfaced this population, which frames what the learned representation adds over conventional detection. Latent dimensionality, training-set size, clustering hyperparameters, and false-positive rates are not specified in the available summary.

Reproducibility considerations

Three items remain unverified in the public material: ablation results comparing the skip-connection autoencoder against a plain autoencoder baseline, quantitative precision-recall on held-out strainmeter channels, and transfer performance on data from non-Parkfield sites. Computational overhead per inference step is likewise unreported, which constrains any conclusion about parameter efficiency. The pipeline — wavelet preprocessing, latent-space compression, unsupervised clustering — is straightforward to port to other continuous deformation datasets or strainmeter arrays. The same paradigm parallels broader shifts in how AI reshapes institutional and operational realities across critical sectors, where high-volume sensor streams dominate the input distribution.