Atlas Deep Geo is a focused technical team developing physics-guided reconstruction methods. We work with operators on real datasets to turn sparse subsurface observations into probabilistic 3D volumes — P10, P50, P90 — with calibrated uncertainty.
Sparse 2D lines. Gaps in a 3D survey. Obscured zones. Each looks different on the surface — but underneath, every one is the same problem: missing observations in a geometry you can map, reconstructed with a learned geological prior. One method handles them all.
The method turns sparse observations into a complete probabilistic volume. A smooth baseline handles the regional trend; a learned residual fills the geological detail; ensemble outputs deliver P10, P50, and P90 — so the uncertainty is part of the answer, not an afterthought.
How it works →Most operators have far more 2D seismic than 3D. Converting that legacy 2D into a usable 3D volume is where the work is focused today. Related challenges — filling gaps in existing 3D surveys, imaging beneath obscured zones, removing acquisition footprint — are problems we take on as the right engagements arise.
The ORCA framework →Two open or independently verifiable datasets. Two problem geometries. The Penobscot study tests 2D-to-3D reconstruction against a well-understood seismic volume with known ground truth. The glacier monitoring study applies the same method to a sparse repeat survey problem. Both were worked through with real partners on real data — the customer received a documented deliverable, and the validated results became the published case study. The uncertainty estimates are reported as they came out, not as we hoped they would.
View studies →Every engagement starts with a specific reconstruction question and a real dataset. We work through it together using ORCA — our physics-guided reconstruction framework that combines a Radial Basis Function (RBF) smooth baseline with a network trained to predict the geological residual the baseline misses. The customer receives a documented deliverable with calibrated P10, P50, and P90 uncertainty envelopes. Where the results hold up and the data can be shared, we publish the case study.
As the same ORCA workflow matures across multiple engagements on the same problem class, customers who use it repeatedly will want to run it themselves. License-style packaging is where this is going — it's just not where it starts.
Most interpolation methods — including Radial Basis Function (RBF) approaches — produce a single best-estimate surface. Sparse inversion does the same. A single answer is useful, but it conceals everything the data cannot resolve.
The ORCA framework takes a different approach. Rather than fitting one surface, ORCA generates an ensemble of geologically plausible reconstructions — each consistent with the observed 2D picks, each honoring the physics of the subsurface. The spread across that ensemble is not noise; it is a direct measure of the irreducible uncertainty that comes from limited spatial sampling. The output is a calibrated P10, P50, and P90 envelope — a range of structures the data is consistent with, not a single answer dressed up as certainty.
That changes the conversation from "here is our best estimate" to "here is what your data actually supports — and here is where it runs out."