Skip to content

EPS 207 · Laboratory in Observational Seismology

This course covers the main tasks of observational seismology — magnitude estimation, phase picking, event location, focal mechanism determination, ground-motion prediction, and seismic tomography — together with the machine learning methods now used for each.

Time and place

  • Tuesdays 9:00–10:59 am, 325 McCone Hall
  • Instructor: Weiqiang Zhu · 285 McCone Hall · zhuwq@berkeley.edu

Schedule

Date Seismology Machine learning
Sep 1 Introduction
Sep 8 Magnitude as a regression, and the b-value of The Geysers Regression & uncertainty
Sep 15 Faulting style at The Geysers Classification & bias-variance
Sep 22 Fault structure from seismicity Clustering
Sep 29 Focal mechanism & moment tensor Inversion I — linear
Oct 6 Location & relocation Inversion II — non-linear
Oct 13 Tomography Inversion III — gradients & backpropagation
Oct 20 Source discrimination Neural networks: training a classifier
Oct 27 Template matching Similarity search & embeddings
Nov 3 Phase picking Neural networks: segmentation & calibration
Nov 10 Denoising Self-supervised learning: denoising
Nov 17 Picking on a dense nodal array Transfer learning & domain shift
Nov 24 Ground-motion prediction Transformers: attention over station sets
Dec 1 Source, path and site spectra Generative models: VAE & flow matching
Dec 8 Final project

What this course is, and how it relates to EPS 130

EPS 130 develops the principal quantities of seismology from the physics: how a wave propagates, how an earthquake is located, what a magnitude measures. EPS 207 takes those same quantities and applies them to real data, in which records are incomplete and every measurement carries a convention and an uncertainty.

This semester we practise on The Geysers, a geothermal field in the Coast Ranges of northern California. It has been producing electricity since 1960, water has been injected into it for decades to sustain the reservoir, and it is among the most seismically active regions in the state as a result — tens of thousands of catalogued earthquakes a year, recorded by a dense local network and archived, along with the rest of northern California, at the Northern California Earthquake Data Center here at Berkeley.

The schedule follows two parallel tracks.

The seismology. The Seismology column names a measurement task — magnitude, picking, location, focal mechanisms, tomography — carried out on real records from this field rather than on a textbook example.

The machine learning. The Machine learning column begins where EPS 130 leaves off: linear regression and uncertainty, followed by classical statistical learning and inverse theory, and then the neural methods. Each is introduced in the context of the seismological problem it addresses.

The first session is an overview of the seismic data available for The Geysers — the catalogue, the waveforms, the arrival times, the injection and production record — so that from the outset you can begin considering which problem you would like to pursue. That material is the project dataset notebook, and every week afterwards returns to it. Open it in Colab: a complete run requires about two gigabytes of memory, and a DataHub session is allocated one. It is also on DataHub, which keeps your edits between visits.

The final project

The project is 60 per cent of the grade; homework is the other 40. You choose the question. The brief sets out nine, none of them with a settled answer: how long the reservoir takes to respond to injection, whether the b-value tracks it, what a machine-learning catalogue adds. Each lists the sessions whose methods apply and a baseline to compare against. You may propose your own instead.

Previous offerings

Fall 2023 is preserved on the fall2023 branch, including the lecture slides.