Listening to Earth's Pulse: Neural Networks Offer a New Approach to Earthquake Forecasting
Researchers at Far Eastern Federal University, working with collaborators from China, have developed a new algorithm for earthquake forecasting. The software produces more accurate models of Earth's crust while requiring relatively modest computing resources. The findings were presented at the conference "Solar–Terrestrial Relations and Physics of Earthquake Precursors."
One established approach to identifying earthquake precursors relies on "listening" to sounds generated within Earth's crust. Periods of heightened tectonic activity are accompanied by distinctive geoacoustic emissions – acoustic signals produced as rocks accumulate immense tectonic stress.
Researchers at Far Eastern Federal University (FEFU), together with colleagues from China, have introduced a technology that sits at the intersection of advanced artificial intelligence and observational geophysics.
They developed a neural network method for analyzing acoustic signals from Earth's crust. The algorithm detects geoacoustic emissions – vibrations generated as stress builds within rock formations. It also uses familiar human-generated vibrations as an additional source of seismic highlighting, helping reveal subsurface structures.
Physics Guides AI Beyond the "Black Box"
The key innovation lies not in the volume of data but in how the data are structured and interpreted. The method combines physics-informed neural networks (PINNs) with the Kolmogorov–Arnold Networks (KAN) architecture. Whereas conventional AI often functions as a "black box" that searches for statistical patterns across massive datasets, the new algorithm incorporates the underlying laws of physics. That allows it to model how seismic waves propagate through rock.
As a result, the mathematical models compensate for limited field observations. They produce detailed two-dimensional and three-dimensional representations of Earth's crust while requiring comparatively modest computational resources.
It would be premature to claim that the neural network is ready to issue precise alerts identifying the exact time and location of future earthquakes. The research remains at the stage of numerical experiments. Russian seismologists also point out that reliable earthquake forecasting still requires much longer observational records and substantially larger high-quality datasets.
That limitation does not diminish the method's practical value. Its primary strength is not predicting the future but providing an exceptionally detailed picture of present-day subsurface structure. High-resolution three-dimensional models of Earth's crust are critically important wherever the cost of geological uncertainty can be measured in billions of rubles or in human lives.
From Oil Exploration to Nuclear Power Plants
Within Russia, the technology aligns closely with the needs of geological exploration. Locating oil, natural gas, mineral deposits, and even kimberlite pipes in Yakutia depends on an accurate understanding of subsurface structure. Russian oil and gas companies are already deploying AI to accelerate geophysical data analysis, and the FEFU approach fits naturally into that broader trend.
A second major application is engineering site investigation. Before constructing nuclear power plants, bridges, tunnels, or hydraulic engineering facilities, engineers must identify concealed faults and subsurface heterogeneity. The algorithm could serve as a digital safeguard that reduces design risks in geologically complex regions.
The Global Race to Forecast Earthquakes
In 2024, an algorithm developed by researchers at the University of Texas and tested in China successfully forecast about 70% of earthquakes one week before they occurred, although it also generated a number of false alarms. The results highlighted both the promise of AI and its current limitations.
By 2025, a scientific consensus had emerged in Russia. Researchers emphasize that no single neural network can reliably forecast earthquakes across the country's diverse geological settings. Algorithms must be carefully adapted to the tectonic conditions of Siberia, the Russian Far East, or the Caucasus. AI is unlikely to replace comprehensive monitoring, but it can make such monitoring substantially more effective.
A Scientific Tool, Not an Oracle
For Russia, with its seismically active regions including the Russian Far East, Kamchatka, the Kuril Islands, and the Baikal Rift Zone, the technology has clear implications for national resilience. Also, the software could prove valuable in earthquake-prone regions across Asia, the Middle East, and Latin America. One practical advantage of the Russian approach is its modest computing requirements. If those low computational demands are confirmed under field conditions, the system could be deployed not only at supercomputing centers but also at remote monitoring stations equipped with basic hardware.
The FEFU project illustrates a mature approach to artificial intelligence. Rather than pursuing the idea of all-powerful neural networks, it combines machine learning with fundamental physics to build hybrid scientific systems.
Over the medium term, the algorithm is likely to become a valuable tool for geological exploration and infrastructure development. It is unlikely, however, to serve as an oracle capable of predicting tectonic disasters with precision, because reliable earthquake forecasting still depends on many years of observations and a comprehensive scientific approach.
