Insights

Reverse-time migration and full-waveform inversion are among the largest commercial high-performance-computing workloads on Earth. An exploration FWI project routinely consumes millions of core-hours, and the cost concentrates in the forward-and-adjoint wave solve: two anisotropic propagations per shot per iteration, times thousands of shots, times tens of iterations, on grids of billions of cells. For thirty years the response was to buy more cluster. The response now is to change the arithmetic of the solve itself. Neural-operator surrogates make a forward call up to three orders of magnitude cheaper. Differentiable simulators return the exact gradient by automatic differentiation, retiring the hand-coded adjoint. Mixed-precision computation and quantized inference fit bigger models on fewer cards. Compute-aware orchestration stops paying for solves that do not move the answer, and sample-efficient methods run fewer of them in the first place. Stacked, these levers do not shave a percentage; they change the regime: velocity building shifts from an overnight batch on a fixed cluster toward an interactive, elastically sized loop, and a solve cheap enough to run a thousand times turns a single best-fit image into a calibrated posterior. The limits are real: surrogates extrapolate silently outside their training prior, learned gradients are approximate, and on a real cluster the bottleneck is frequently I/O rather than FLOPs. The discipline the stack demands is the one it rewards: gate the surrogate, keep an exact reference in the loop, and measure the whole pipeline.

Why the well count and its intake gate were the real project-management artifact of a three-phase carbonate borehole-imaging program in Oman. The running tally read 5-in / 0-kept, then 10-in / 8-kept, then 25-asked / 8-delivered, and we tracked the asked-versus-delivered gap every month on a data-management dashboard. This is a first-person account of the ledger discipline itself: how we ran the program on one number so that, by the final report, no one was surprised by the 8.

The slowest, most expensive part of a carbon-storage programme is not injecting the CO2, it is deciding where to even look. An agentic screening pipeline gives each of the five geological suitability criteria its own software agent, reconciles them into one auditable GO / REVIEW / NO-GO verdict, and compresses a first-pass screening cycle from roughly four months per site to under ten minutes in a labeled synthetic demonstration, so a whole licence portfolio can be triaged in an afternoon.

The engagement we were funded to run in August 2020 was a 10-month geomechanics ANN estimating in-situ stress and elastic moduli from borehole deformation. What shipped, three years later, was a fracture, bedding and vug detection transformer. This is the honest account of how the technical core pivoted without the contract dying, and why the phase-gate structure is what made that survivable.

H-κ stacking assumes a flat horizontal Moho in its moveout equations. Tilt the interface by ten degrees and the crustal-thickness estimate can move by about two kilometres and Vp/Vs by a few hundredths, enough to misread crustal composition, while the stack itself looks perfectly healthy. The error is a systematic bias, not scatter, so more events make it worse. What the practitioner should check, and how to represent dip instead of assuming it away.

One craton, roughly ten thousand receiver functions, three years of expert picking, and still no product. The gap between a beautiful single-region research result and a tool that estimates crustal thickness anywhere is a five-stage funnel of curation, generalisation, honest validation, and packaging. Most geoscience ML projects stall in it. Here is how the funnel works, why the last stage pays for the first, and why the next region should cost an API call rather than another PhD.