Sibirsky tract: neural network to build roads in permafrost conditions

Scientists from Perm have trained AI to predict the behavior of road structures reinforced with geosynthetic materials on weak soils — swampy soils and permafrost conditions. The development will help to build reliable all-season roads where fragile winter roads now have to be used, and provides forecasts with an accuracy of 92%. In difficult climatic zones, conventional logging routes are rapidly destroyed, and existing calculation methods do not allow accurate prediction of how new materials will behave. Experts consider domestic development promising for the introduction of AI in road structures.

The neural network determines the most important indicators of the roadway

Employees of Perm National Research Polytechnic University (PNRPU) have created a neural network that accurately predicts how a geosynthetic road will behave. There are no analogues of such a system anywhere in the world yet, scientists assure. First, they conducted hundreds of computer experiments, simulating different variants of roads and soils and recording the subsidence of the coating and stress in the reinforcing material. Then, a neural network was trained on this data, which, according to 13 input design parameters, produces two key results — the drawdown value and the voltage level.

Photo: RIA Novosti/Ilya Naimushin

— The test results confirmed the high efficiency of the developed neural network. The model demonstrated stable forecasting accuracy at the level of 90.76% with an error of less than 10%, which fully meets the requirements for practical use in road construction. The tests carried out showed the reliability of the system — during independent launches, the accuracy of forecasts remained in the range from 88.27 to 92.06%," said Vladimir Kleveko, Associate Professor of the Department of Highways and Bridges at PNRPU, Candidate of Technical Sciences.

In addition, the neural network shows which parameters most affect the strength of the road. It turned out that the key factor is considered to be the strength of the soil — it primarily depends on it how quickly the track forms and what kind of load will fall on the geosynthetic grid. To prevent this problem, the thickness of the asphalt surface is most important, and for the durability of the mesh itself, its rigidity is most important.

Photo: RIA Novosti/Ilya Naimushin

In the future, scientists plan to patent software based on the new technique, which will allow it to be implemented in real-world design. This technique speeds up calculations and does not require expensive software, so designers and forestry companies will have an affordable tool for creating reliable year-round roads on weak soils. It is expected that this will reduce the cost of construction, make the development of remote forests more profitable, and reducing logistical costs will affect the cost of all forest products, from building materials to paper and furniture.

AI without borders: Chinese models capture the US and help Russia

Asian neural networks are changing the rules of the game in the global AI systems market

The problem of roads in Russia has not been solved yet.

The Russian forestry industry, which owns a huge part of the world's timber reserves, is experiencing a crisis — although revenues are growing, the volume of logging in the first ten months of 2025 fell by 9% due to logistical problems. Most of the difficulties are in remote regions where there are few roads, and natural conditions (swamps, rivers, permafrost) greatly complicate the construction of normal infrastructure. At the same time, the richest forests are located exactly where it is impossible to reach, and building roads in such places is too expensive and always does not pay off, explained Vladimir Kleveko.

Photo: TASS/Yamal-Media/Fedor Voronov

The PNRPU specialist clarified that for a long time the industry relied on "winter roads" — temporary roads on frozen ground and ice, which existed only in winter and allowed reaching the most remote forest areas. But due to climate change, frosts are less severe, winter has decreased by about 15-20%, thaws often occur, and permafrost is melting in the northern regions, so winter camps have become unpredictable and unsafe.

Now it is necessary to build year-round roads, but weak swampy soils cannot withstand heavy machinery: deep ruts form, roads deteriorate quickly, and transportation becomes more expensive. To strengthen such roads, geosynthetics are used — durable synthetic materials that are placed between the layers of the road to distribute the load and prevent the rubble from falling through.

Photo: RIA Novosti/Ilya Naimushin

— This makes the roads more stable even on difficult ground. The problem is that the old methods of calculating such roads are no longer suitable.: they inaccurately predict how new materials will behave on weak soils, so the engineer either makes the project unreasonably expensive, or risks getting a road that will quickly collapse," the expert explained.

To solve these problems, a new neural network was developed, its creators noted.

A mistake has been made: new software will help neural networks fix their own bugs

How will the development of self-regulating intelligence help ordinary users get into programming

The practical potential of the development surprised the experts

The technology solves one of Russia's key problems — the transport accessibility of remote and resource-intensive regions. And predicting with 92% accuracy is more useful than it seems at first glance. However, it will have a negative economic effect if the cost of implementing and maintaining the technology exceeds the savings on materials and engineering and geological surveys, said Ruslan Dolgopolov, head of the Gazprom ID Operator product group.

"The construction of reliable all-season roads will open up access to resources that are currently unprofitable to develop due to logistical problems," he added.

Photo: RIA Novosti/Evgeniya Novozhenina

Previously, the technology of using synthetic components in asphalt chips to slow down wear was based solely on mathematical modeling. Innovation lies in the use of neural networks. Multilayer networks allow you to take into account a huge number of parameters and simulate the behavior of complex materials with greater speed, which makes the process much more efficient, says Sergey Golitsyn, head of T1 AI.

— There is potential for technology development in the following areas: reducing the cost of construction and logistics in the forestry industry, developing remote areas with rich resources, and replacing unreliable winter roads with cost—effective all-season roads," explained Ilya Revin, PhD, Senior researcher at the ITMO Institute of AI, expert at the National Center for Cognitive Research at ITMO.

He noted that the method has already proved its practical applicability and can become the foundation for the sustainable development of the forestry industry in difficult conditions.