A growing field of research is changing how AI is applied to subsurface problems. Instead of asking a neural network to learn everything from data alone, physics-informed AI constrains the solution space using what we already know — the laws of wave propagation, the geometry of acquisition, the character of geology. Atlas Deep Geo builds at that intersection.
Conventional AI asks a network to learn everything from data alone — an approach that works when examples are plentiful and the problem is well-defined. Subsurface reconstruction is neither. Data is sparse, ground truth is rarely available, and the space of possible models is enormous. Without constraints, a network will find a model that fits the observations, but it has no reason to prefer one that is geologically realistic over one that simply happens to match the numbers.
Physics-informed AI changes the framing by embedding what we already know directly into how the model learns. Wave propagation principles, acquisition geometry, and the statistical character of geology constrain the solution space before the network ever sees the data — so rather than searching blindly, the AI is guided toward answers that are physically plausible from the start. The network still learns, but it learns within bounds that reflect how the subsurface actually behaves.
That constraint is what makes the output trustworthy. Because the physics narrows the solution space without collapsing it to a single answer, the result is not a point estimate but a calibrated distribution — a range of geologically plausible structures consistent with the observations, with the spread across that range quantifying exactly what the data leaves unresolved. That is a fundamentally different and more honest deliverable than a single best-guess surface.
The laws of wave propagation, acquisition geometry, and subsurface geology constrain what solutions are physically plausible. The network learns within those bounds — not in spite of them.
Start from the simplest model that fits the data. Add geological detail only as the observations require it. Never invent complexity the data doesn't support.
A single best-estimate answer conceals what the data cannot resolve. Physics-informed AI delivers a calibrated range — P10, P50, P90 — so the uncertainty is part of the result, not an afterthought.
The application of physics-informed AI to subsurface problems has been shaped by a handful of research groups over the past two decades. The people behind Atlas Deep Geo are among them — with careers that span the development of compressive sensing reconstruction, sparse transform methods, and uncertainty quantification workflows that have been validated at scale across global deployments. The research community referenced here is one we have worked alongside and contributed to for decades — not one we are catching up to.
One of the foundational research programs in seismic imaging and reconstruction, with deep work in physics-constrained learning and subsurface data science.
sep.stanford.edu →Seismic Laboratory for Imaging and Modeling. Felix Herrmann's group — leading research in physics-informed AI for seismic data, full-waveform inversion, and uncertainty quantification at scale.
slim.gatech.edu →Seismic Wave Analysis Group at King Abdullah University of Science and Technology — advancing physics-informed AI for geophysics in carbonate and arid basin settings.
swagroup.kaust.edu.sa →ORCA is how Atlas Deep Geo applies physics-informed AI to the specific problem of seismic reconstruction. It is not a general-purpose AI tool — it is a focused implementation of the principles above, built for one problem class and validated against real data.
Turns sparse 2D seismic observations into probabilistic 3D volumes. A physics-grounded smooth baseline handles the regional trend; a network trained on geologically realistic synthetics corrects the detail the baseline misses. The output is a calibrated P10, P50, and P90 envelope — not a single answer.
The network never operates without physics. Acquisition geometry, wave propagation constraints, and geological prior knowledge bound every reconstruction it produces.
Read the full ORCA framework →Most reconstruction methods produce one answer. ORCA produces the simplest ensemble of geologically plausible answers consistent with the data — and reports the spread across that ensemble as calibrated uncertainty.
That means every deliverable comes with an honest statement of what the data constrains and what it does not. The uncertainty shrinks where observations are dense and widens where they are sparse — exactly as it should.
See the Penobscot case study →