Freeze-thaw Damage Model for Compressive Strength of Flowable Solidified Soil Based on Machine Learning

Imagine you are building a sandcastle, but instead of just sand, you are mixing in special glue and water to make it super strong. This is a bit like what engineers do with "flowable solidified soil." They take soft, squishy dirt (sometimes even waste from factories) and mix it with binders to create a solid material perfect for filling holes under roads or building foundations. It's a green, smart way to reuse waste and build strong structures. But there's a catch: if you build this in a cold place, the weather plays a nasty game of "freeze and thaw." When water inside the soil freezes, it expands like a balloon, cracking the material from the inside. When it melts, the cracks stay. Do this over and over, and your strong sandcastle turns into a pile of crumbs.

For a long time, predicting exactly how much strength this soil would lose after hundreds of these freeze-thaw cycles was like trying to guess the weather a month in advance using only a barometer. Engineers used old formulas that were okay, but they often missed the mark because the relationship between the mix ingredients and the damage is messy and complicated. That's where the paper you're about to read comes in. It swaps the old barometer for a high-tech crystal ball called "Machine Learning." Instead of simple math, it uses computer algorithms that learn from real experiments to spot patterns humans might miss. The big question is: Can these smart computers predict exactly how strong the soil will be after the cold weather beats it up, and which ingredient matters the most?

The Digital Crystal Ball: Predicting Frozen Soil Strength

In this study, a team of researchers from Zhengzhou University and partners in Henan, China, decided to teach computers how to predict the fate of flowable solidified soil in freezing conditions. They didn't just guess; they ran a massive experiment. They took silty clay (a type of soil) and mixed it with different amounts of "solidifier" (the glue), "admixture" (a helper chemical), and water. They created 50 different batches of this soil, each with a unique recipe. Then, they subjected these little cubes of soil to a brutal cycle: freezing them to -10 ℃ and thawing them to 20 ℃, repeating this process up to 50 times. After every round of freezing and thawing, they crushed the cubes to measure their compressive strength.

With this mountain of data, the researchers built four different "digital brains" (machine learning models) to learn the rules of the game:

1. Decision Tree (DT): Like a flowchart that asks simple yes/no questions.
2. BP Neural Network (BP): A model that mimics how human neurons connect.
3. Support Vector Regression (SVR): A method that tries to find the best line to fit the data points.
4. XGBoost: An advanced team of decision trees working together to make a smarter prediction.

The goal was to see which "brain" could best predict the soil's strength after the freeze-thaw torture, based on just four inputs: how much glue was used, how much helper chemical, the water-to-soil ratio, and how many times it was frozen.

The Showdown: Who Got the Best Grades?

When the models were tested, the results were a mix of "wow" and "oops."

The Decision Tree (DT) model was a bit of a show-off in the classroom (the training set). It memorized the answers perfectly, getting a near-perfect score. But when it faced a new test (the test set), it completely fell apart. It suffered from "overfitting," which is like a student who memorized the practice exam but panicked when the real questions were slightly different. Its predictions were all over the place, with a low accuracy score of 0.79.

The BP Neural Network and SVR models were much steadier. They didn't memorize the answers as obsessively, so they handled the new test well, achieving a high accuracy of 0.98. They were reliable, but they still had a few wobbly moments.

The real star of the show was XGBoost. This ensemble model, which combines the wisdom of many smaller trees, struck the perfect balance. It didn't just memorize; it understood the patterns. It achieved a test accuracy of 0.95, which was slightly lower than the top two but came with a major bonus: it was the most stable. When the researchers looked at the errors (how far off the predictions were), XGBoost's mistakes were small and consistent, whereas the others had wild swings. It was the most robust "digital brain" for this job.

The Secret Ingredient: What Actually Matters?

Once they had their best model (XGBoost), the researchers used a special tool called SHAP to ask the computer: "Which factor is actually causing the soil to break?"

The computer gave a clear ranking of importance:

1. Freeze-Thaw Cycles (F-T): This was the undisputed boss. The number of times the soil was frozen and thawed had the biggest impact on its strength. It was the dominant factor.
2. Solidifier Dosage (SD): How much glue you put in was the second most important thing.
3. Admixture Dosage (AD): The helper chemical mattered, but less than the glue.
4. Water-Solid Ratio (WS): The amount of water was the least influential of the four, though it still played a role.

The analysis also showed how these factors worked. For example, more freeze-thaw cycles (a high number) always pushed the predicted strength down. More solidifier (glue) pushed the strength up. But the relationship wasn't always a straight line; sometimes having too much or too little of a specific ingredient had different effects, and the machine learning model was smart enough to catch those subtle twists.

The Takeaway

This paper doesn't just say "computers are cool." It proves that for predicting how flowable solidified soil survives in cold climates, a smart machine learning model (specifically XGBoost) is far superior to old-school guessing or simple formulas. It can predict the damage with high accuracy and, more importantly, it tells engineers exactly which variables to watch.

The researchers found that while the soil's recipe matters, the sheer number of freeze-thaw cycles is the main villain. By using this model, engineers in cold regions can now quickly estimate how long their soil structures will last without having to wait years for real-world weather to test them. It's a faster, smarter way to build safer roads and foundations, ensuring that the "sandcastles" of the future don't crumble when the winter comes.