Numerical homogenization of a heat transfer problem with phase transitions in heterogeneous frozen soils

In the vast, frozen north, the ground is not merely dirt; it is a structural foundation that holds up the world's largest cities, pipelines, and roads. For centuries, this permafrost has remained solid, its frozen state acting as a rigid bedrock for human construction. However, as the climate warms, this frozen ground is beginning to thaw. When the ice within the soil melts, the ground loses its strength, leading to uneven sinking and the cracking of buildings. Predicting exactly how deep this thaw will go is a matter of safety and engineering, but the soil itself presents a massive complication. Real soil is not a uniform block of ice; it is a chaotic mixture of layers, lenses, and pockets of different materials, with properties that change over distances as small as a few centimeters. To simulate the melting of such complex ground with perfect accuracy, a computer would need to track every single layer, a task that requires more computing power than is currently available for large-scale, long-term forecasts.

To solve this, researchers turn to a technique called numerical homogenization. Imagine trying to predict the flow of water through a sponge. Instead of modeling every tiny hole and fiber, you treat the sponge as a single material with an "average" ability to let water pass through. This is the essence of homogenization: replacing a messy, detailed reality with a simpler, averaged model that is fast enough to run on a computer. However, when the ground is melting, this averaging becomes a trap. The boundary between frozen and thawed soil—the "front"—does not stop neatly at the edge of a computer's calculation block. It cuts right through the middle. This creates a dilemma: if a block of soil is half-frozen and half-thawed, how should the computer decide its properties? Should it act more like ice, more like water, or something in between? The choice made here determines whether the model predicts a building will sink or stand firm.

In a recent study, Timur Zakharov from North-Eastern Federal University tackled this specific dilemma by testing three different ways to handle these "mixed" blocks of soil. The research focused on a heated building sitting on frozen ground, a scenario where the warmth from the structure slowly melts the soil beneath it. The team created a digital twin of this environment, using a realistic, random pattern of soil conductivity to mimic the natural chaos of the ground. They then ran the simulation on a coarse grid, where each block of soil was much larger than the actual layers of dirt and ice, forcing the phase-change front to slice through the middle of these blocks. They compared three distinct methods for deciding what happens inside those sliced blocks.

The first method, known as the table approach, pre-calculates a list of properties for every possible state of a soil block and looks up the answer based on the block's average temperature. The second method, a nodal scheme, decides the state of the soil at the corners of the blocks and averages those decisions. The third method, called h-smoothing, calculates the exact fraction of the block that is melted based on a linear temperature gradient, essentially measuring the slice of the block that has turned to water. The researchers ran these simulations in both two and three dimensions, comparing the results against a highly detailed "reference" solution that tracked every layer of soil.

The findings revealed a surprising trade-off that challenges how engineers might choose their tools. When the goal is to predict the temperature of the soil, the h-smoothing method proved to be the most accurate. It tracked the heat flow with the highest precision, especially on the coarse grids where the other methods struggled. However, when the researchers looked at the position of the melting front—the actual boundary between frozen and thawed ground—the story changed completely. The h-smoothing method, while good at temperature, consistently overestimated how deep the thaw would go. It predicted the ground would melt further than it actually did.

The reason for this discrepancy lies in the laws of energy conservation. The h-smoothing method, in its effort to be precise about the shape of the melting zone, inadvertently added a small amount of extra heat to the system with each time step. Over the course of a year, this tiny surplus accumulated, causing the simulated ground to melt more than the real ground would. In contrast, the other two methods, particularly the table approach and the nodal scheme, were "conservative." They did not add or lose energy, and as a result, they predicted the depth of the thaw much more accurately, even if their temperature readings were slightly less precise. The study showed that if a researcher needs to know the exact temperature of the soil, h-smoothing is the superior choice. But if the goal is to know how deep the foundation will sink due to thawing, the conservative methods are far more reliable.

The research also explored what happens when the temperature at which soil melts is not the same everywhere, but varies across different layers. In this more complex scenario, the table-based method, which uses a detailed lookup of properties, outperformed the simpler two-tensor methods in predicting the position of the thaw front. This suggests that for highly variable ground conditions, a more detailed pre-computed map of soil states is necessary. Ultimately, the study provides a clear guide for engineers and scientists: there is no single "best" method. The choice depends entirely on what question is being asked. If the priority is thermal accuracy, one method wins; if the priority is the physical boundary of the thaw, another takes the lead. By clarifying these differences, the work helps ensure that the models used to protect buildings in the Arctic are not just fast, but also trustworthy.