Gradient Boosting-Based Computational Method for Predicting Uniaxial Compressive Strength of Cemented Hydraulic Backfill

Deep beneath the earth's surface, where miners extract valuable resources, vast empty spaces are left behind. If these voids are not supported, the ground above can collapse, creating a dangerous environment for workers and threatening the stability of the entire operation. To prevent this, engineers fill these empty chambers with a special mixture known as cemented hydraulic backfill. This material is essentially a slurry made from crushed waste rocks, tailings from the mining process, water, and cement. Once pumped into the underground voids, it hardens over time, acting as a solid floor that supports the mine roof and allows for safer, more efficient extraction of remaining minerals. The most critical measure of how well this backfill will perform is its ability to resist being crushed under pressure, a property engineers call uniaxial compressive strength. Predicting this strength accurately is vital; if the mixture is too weak, the mine could fail, but if it is over-engineered, the project becomes unnecessarily expensive.

For decades, determining the right mix of ingredients has relied on trial and error or simple formulas that often miss the complex, non-linear ways these materials interact. A new study published by researchers from the China University of Mining and Technology and international partners offers a different approach. Instead of relying on traditional equations, they turned to a family of advanced computer algorithms known as gradient boosting machines. These are powerful tools that learn from data by building a series of simple models, each one correcting the mistakes of the last, to create a highly accurate predictor. The researchers gathered a massive dataset of 1,641 real-world examples, recording the specific amounts of cement used, the duration of the curing time, the concentration of solids in the mix, and the ratio of aggregate to tailings. They then fed this information into four different types of these learning algorithms to see which one could best predict the final strength of the hardened backfill.

The study tested several variations of these learning models, including LightGBM, CatBoost, AdaBoost, and a standard Gradient Boosting Machine. The goal was not just to find a model that could guess the strength, but to find one that could do so reliably on new, unseen data without getting confused by the noise in the information. After training the models on 80 percent of the data and testing them on the remaining 20 percent, the results were clear. The standard Gradient Boosting Machine model emerged as the most effective tool. It achieved a high level of accuracy, correctly predicting the strength values with a statistical score of 0.86 on the test data, while keeping the average error very low. This performance was superior to the other models, which either struggled to generalize to new data or produced less consistent results. The findings suggest that this specific type of algorithm is particularly well-suited for capturing the intricate relationships between the ingredients and the final strength of the backfill.

Beyond simply making a prediction, the researchers wanted to understand why the models made the decisions they did. In many advanced computer systems, the inner workings are a "black box," meaning even the creators cannot easily explain how an input leads to an output. To solve this, the team used a technique called SHAP analysis, which acts like a spotlight, revealing exactly how much each ingredient contributed to the final result. The analysis showed that two factors dominated the outcome: the amount of cement in the mix and the time allowed for the material to cure. These two variables had the most significant positive impact on the strength. In contrast, the concentration of solids and the ratio of aggregate to tailings had a much smaller influence on the final result. This insight is crucial for engineers, as it tells them exactly where to focus their attention when designing a mix to ensure safety and stability.

The study concludes that using these explainable machine learning models provides a reliable and transparent way to design cemented hydraulic backfill. By accurately predicting how strong the material will be before it is even poured, mining operations can optimize their mix designs, potentially saving money on materials while ensuring the underground environment remains safe. The researchers emphasize that while their model is a significant step forward, it is most effective when used as a decision-support tool alongside traditional engineering judgment. They suggest that future work could involve testing these models against real-world field measurements to further refine their accuracy. Ultimately, this research offers a clearer path to safer underground mining, turning complex data into practical guidance for the engineers who keep the mines stable.