MIPT: Algorithm for converting seismic waves into a detailed map of underground structures

2025: Introduction of the algorithm for converting seismic waves into a detailed map of underground structures

developed Algorithm by scientists, MIPT in just dozens of mathematical operations (instead of thousands) converts seismic waves into a detailed map of underground structures. The system will help to quickly search for minerals, draw maps for study Lands and assess seismic risks at. construction The results of the study are presented in the archive of preprints of scientific articles arxiv.org, and the original code is available to scientists around the world on the platform. The GitHub university announced this on September 16, 2025.

𝓲

Фото: Celes.club

To look into the bowels of the Earth, scientists use principles similar to echolocation - they send elastic vibrations deep into tens of kilometers. By the nature of how fluctuations are reflected from different rocks, a map of the location of the main structures in the studied media is built.

Diffusion models are used for this task: to restore a clear picture of the bowels of the Earth, they require hundreds or thousands of calls to the neural network.

This cI2SB method (conditional Image-to-Image Schrödinger Bridge) speeds up the process several times.

It is based on the diffusion equation, which solves the well-known problem of the Schrödinger Bridge, which scientists have adapted to work with velocity maps.

Unlike diffusion models, the system does not recover velocity maps from random noise, but builds a "bridge" between two specific points: the approximate (blurred) velocity model and the desired - detailed reference model.

This reduces the required number of calculations by several times - up to 50-100 calls with comparable quality, which significantly speeds up the process.

𝓲

Фото: Celes.club

{{quote 'It is as if you saw underground structures through a foggy window, but step by step gradually erased this fog, revealing a clear picture of what is hidden underground. So is our system: at the entrance we have a blurred model and raw recording of seismic waves. With the help of a neural network on the U-Net architecture, which we trained as part of the cI2SB approach, the system "refines" the map step by step and adds details to it. The output is a detailed velocity map, which corresponds to a realistic geological structure from the training sample, - said Andrei Stankevich, graduate student, assistant at the Department of Computer Science and Computational Mathematics of the Moscow Institute of Physics and Technology. }}

The method has already been tested on large sets of synthetic data (OpenFWI - Vel, Fault, Style). As a result, the cI2SB restores velocity maps more accurately than previous diffusion models, does it 10-20 times faster and at the same time retains subtle geological features.

But there are also limitations. If the actual input data are very different from those on which the model was trained (other equipment, geological conditions, basins, etc.), the quality of the model may suffer.

Making the method more stable and bringing it closer to processing real seismic data is the next task of scientists. Also in the future, the algorithm can be adapted to handle natural earthquakes. To do this, you need to take into account additional types of vibrations, as well as uncontrolled sources of waves.