Fig. 9
Earthquake early warning pipeline. (a) Earthquake early warning workflow: an on-site sensor monitors for any live ground motion. After P-wave detection, a three-second window of ground motion is used to compute a vector of intensity measures (IMs). The IM vector is fed to a pre-trained feed-forward neural network (FFNN) to estimate the corresponding point in the hierarchical variational autoencoder (HVAE) latent space. The estimated point is then decoded into the full response spectrum and used to conduct the final early warning. (b)-(c) True latent variables plotted against predicted latent variables from the FFNN, (b) \(z_0\) (\(R^2=0.907\) for all data) and (c) \(z_1\) (\(R^2=0.908\) for all data).
