A geophysics primer for ML readers

This is the closing section of Part 0 and the only one that does not involve neural networks. Its job is to make sure that when Part 1 says "the wave equation", Part 5 says "Marmousi", and Part 10 says "AVO from PINN-augmented FWI", you have a one-paragraph anchor for every one of those terms. We assumed you might bring zero ML knowledge to Part 0; we assumed you might bring only modest geophysics literacy. Parts 1–10 will be using both vocabularies heavily. This section is the seismic-side dictionary.

How to use this card

Read it now if you have a geophysics background — you will skim it in two minutes and confirm there are no surprises in our terminology. If you do not, read it now too — you will see twenty-two concepts you will be expected to recognise. Either way, bookmark this section and come back to it any time a Part-N lesson mentions a term you cannot place. Cards with a clickable title cross-reference into the relevant section of this textbook, of the Acquisition textbook on the same site, or of the Processing textbook — wherever the topic is treated in full.

What this primer does NOT cover

Real exploration seismology has hundreds of vocabulary items. This primer covers the twenty-two that will actually appear in Parts 1–10 of this textbook. If you find a term in the wider literature that we did not list — deconvolution, demultiple, common-offset gather, ghost notch, AVA, EI — the Acquisition or Processing textbook on this site has it, and we will name those texts whenever a deeper treatment is warranted.

What you should know if you are coming from a pure-ML background

Three meta-insights specifically for the ML reader:

• The forward model is a PDE solve. In standard supervised learning, the forward model is whatever the dataset assumed; you do not solve a differential equation. In seismic forward modelling, you solve the wave (or eikonal, or Helmholtz) equation. PINN-FWI replaces the classical numerical solver with a neural network, and it is that change that makes the data-flow end-to-end-differentiable.
• The "labels" are noisy and indirect. The "y" we want to predict (the velocity model) is never directly observable. We see only the wavefield it produced, contaminated by source ghosts, multiples, attenuation, anisotropy, and noise. Inversion is the art of recovering the cause from the effect, with a PDE in between.
• The "input" is also a parameter. In standard supervised learning, x is fixed and we learn weights. In PINN-FWI, both the velocity model and the network weights are simultaneously updated. The network plays a dual role: it is the wavefield representation and the velocity model representation. We will untangle these roles in Parts 1, 5, and 6.

Expertise checkpoint - end of Part 0

If you started Part 0 with no neural-network knowledge, you should now be able to:

• Hand-trace a forward + backward pass through a 3-neuron MLP and explain every intermediate value.
• Pick an activation function for a given problem class on inductive-bias grounds, and defend the choice.
• Read a 2D loss landscape and identify the minimum, valleys, and saddles by eye.
• Set up a gradient-descent run with a sensible learning rate, momentum, and Adam configuration.
• Predict whether a vanilla MLP will fail on a high-frequency target, and name two architectural fixes (Fourier features, SIREN).
• Define every term in the geophysics primer above, in your own words.

If any of those is shaky, the relevant section is one click away in the sidebar. If all are solid, you are ready for Part 1, where we put the pieces together to formulate the first PINN — the loss function that includes a physics-based PDE residual.

References

• Sheriff, R.E., Geldart, L.P. (1995). Exploration Seismology (2nd ed.). Cambridge University Press.
• Yilmaz, Ö. (2001). Seismic Data Analysis (2 vols.). Society of Exploration Geophysicists.
• Aki, K., Richards, P.G. (2002). Quantitative Seismology (2nd ed.). University Science Books.
• Virieux, J., Operto, S. (2009). An overview of full-waveform inversion in exploration geophysics. Geophysics 74(6), WCC1–WCC26.