NO. 43 · Computational & Data

Machine Learning in Seismic Practice

Not a demo reel. Where learned methods actually ship across the seismic value chain: compressed sensing and deblending in acquisition, denoising and reconstruction in processing, facies in interpretation, and the discipline that keeps all of it honest.

You can say where ML earns a place in a seismic workflow and where it still loses to physics, run the flagship applications from survey design to QI, and hold any learned result to the regularization and calibration standards a client audit would.

12 competencies · 6 interactive widget challenges · 6 to 9 hours of guided study

For working geophysicists deciding where ML belongs in the flow

Draws on: 3D Seismic Interpretation Seismic Acquisition Seismic Data Processing Statistics and Data Science for Researchers

The scorecard

Where ML fits, and where it does not

Every task in the processing sequence has a physics baseline; the scorecard is the habit of asking what the network must beat before you train it.

Reads: ML in processing: where it fits

Smart acquisition

Compressed sensing and optimal samplingwidget challenge

Recovering the wavefield you deliberately did not record is the boldest data-generation trick in geophysics, and it rests on sparsity you can test.

Reads: Compressed sensing fundamentals Nonuniform optimal sampling (NOS)

Blended shooting and deblendingwidget challenge

Firing sources on top of each other buys survey time; iterative deblending is the computational bill, and it usually clears.

Reads: Blended acquisition & iterative deblending

ML-guided survey designwidget challenge

A learned model proposing geometries closes the loop between imaging objective and layout; judge its proposals with the design rules you already own.

Reads: ML-guided survey design

Passive, ambient, and microseismic

Noise recorded for free is data when correlation and location algorithms are good enough; passive methods are acquisition's quiet computational frontier.

Reads: Passive & ambient-noise acquisition Microseismic monitoring

Learned processing

Denoising with CNNswidget challenge

The first learned processor most shops adopt; the question is never can it denoise, it is what signal it quietly ate.

Reads: Denoising with CNNs

Interpolation and reconstructionwidget challenge

Generating the traces you never recorded fills the gaps acquisition left; the honest question is where interpolation becomes invention.

Reads: Interpolation / reconstruction

ML first-break pickingwidget challenge

Thousands of first breaks feed statics and tomography; a picker that is fast and mostly right beats one that is perfect and never finished.

Reads: ML-assisted first-break picking

ML inside FWI

Learned gradients and surrogates cut the cost of the most expensive loop in geophysics; the physics still grades the answer.

Reads: ML for FWI gradient acceleration

Learned interpretation

Machine-learning QI

Facies classification at cube scale is where learned interpretation pays; calibration decides whether its confidence means anything.

Reads: Machine-learning QI: neural networks for facies classification

The discipline

Regularization and calibration

Ridge and lasso keep a model from memorizing the survey, and calibration makes its probabilities auditable; together they are the difference between shipped and shelved.

Reads: Regularisation: ridge, lasso, elastic net Calibration and probability outputs

When ML loses to physics

The senior skill: naming the tasks where a wave equation or a lookup table still beats any network, and declining gracefully.

Reads: When NOT to use ML