A large open multi-source soil compaction corpus and pretrained tabular estimators from index properties

Every road, dam, and building foundation begins with a simple, critical question: how tightly can we pack the earth beneath it? When engineers build with soil, they do not just pile it up; they compact it, squeezing the particles together to remove air and water, creating a solid base that will not sink or shift under weight. To do this safely, they must know two specific numbers for every type of dirt they use. First, they need to know the maximum dry density, which is the heaviest the soil can possibly become when squeezed. Second, they need the optimum moisture content, the exact amount of water the soil needs to reach that maximum weight. If the soil is too dry, the particles rub against each other and refuse to pack tight. If it is too wet, the water acts as a lubricant that pushes the particles apart. Finding the sweet spot requires a physical test where engineers mix soil with different amounts of water, hammer it into a mold, and weigh it. This process is slow, expensive, and must be repeated for every new pile of dirt a project encounters, often taking a full day of lab work just to get one answer.

For decades, researchers have tried to skip this slow process by guessing the answer based on simpler, faster tests. They hoped that by measuring basic properties of the soil—like how much clay it contains or how sticky it is when wet—they could predict the heavy-compaction numbers without doing the full test. However, these previous attempts have been limited. Most studies relied on small collections of data from just one laboratory or one type of soil, meaning their "rules" often failed when applied to different dirt found in different places. Furthermore, many of these studies treated missing information as a problem to be fixed by inventing a number, rather than accepting that some soils simply do not have certain properties.

A team of researchers at King Mongkut's University of Technology Thonburi has now tackled this problem by building something much larger and more honest: a massive, open collection of soil data drawn from dozens of different studies and laboratories around the world. Instead of guessing or inventing data, they gathered nearly 5,000 records of actual soil tests, carefully checking each one to ensure it followed the laws of physics. They found that nearly twelve percent of the published data they collected described soil states that were physically impossible, such as soil that was more than one hundred percent saturated with water, a condition that cannot exist. By removing these impossible records, they created a clean, reliable database that spans everything from clean sand to heavy clay, covering four different levels of compaction energy.

With this new database in hand, the researchers tested a modern approach to prediction using a type of artificial intelligence known as a tabular foundation model. Unlike traditional computer programs that learn by adjusting their own internal settings based on the specific data they are given, this model was pre-trained on millions of synthetic examples and then kept frozen. It does not "learn" the soil data in the traditional sense; instead, it looks at the new soil records as context, much like a student reading a textbook example before solving a problem. This method allowed the model to predict the maximum dry density and optimum moisture content with high accuracy, achieving a level of precision that rivals the physical tests themselves. The model successfully predicted the maximum dry density with an error of only 0.067 Mg/m³ and the moisture content with an error of 1.67 percentage points.

The study also revealed exactly which pieces of information matter most. When the researchers removed the data describing the size of the soil particles, the accuracy of the predictions dropped significantly, proving that knowing whether the soil is sandy or silty is the single most important factor. They also found that the consistency limits, which measure how the soil behaves when wet or dry, were crucial for fine-grained soils but irrelevant for clean sands. Crucially, the model learned to handle missing information correctly. When a soil sample had no plastic limit because it was a clean sand, the model recognized this absence as a specific state rather than trying to fill it with a fake number. This allowed the system to work across the entire spectrum of soil types, from the coarsest gravel to the finest clay, without breaking down.

However, the researchers were careful to note the limits of their work. While the model performed exceptionally well at predicting results for soils similar to those in their database, it has not yet been tested on completely new types of soil from unseen sources or different laboratories. The study also highlighted a trade-off between pure accuracy and physical safety. The most accurate model occasionally predicted a combination of density and moisture that would imply the soil was more than one hundred percent saturated, a physical impossibility. To solve this, the team also derived a simpler, mathematical formula that guarantees the results will always be physically possible, though this formula is slightly less accurate than the complex computer model.

Ultimately, this work provides a new foundation for the future of geotechnical engineering. By releasing this large, verified dataset and demonstrating how modern artificial intelligence can learn from it without overfitting, the researchers have given the scientific community a shared benchmark. This allows future engineers to develop better tools for predicting soil behavior, potentially saving time and money on construction projects while ensuring that the ground beneath our buildings remains safe and stable. The study confirms that while we cannot yet predict the behavior of every unknown soil perfectly, we have moved far beyond the era of guessing based on small, isolated experiments. We now have a clear, data-driven path forward for understanding how the earth behaves under pressure.