📄 earth science

Physics-Residual Learning with Few-Shot Calibration for Cross-Geological UCS Prediction: A Model-Agnostic Framework

This paper introduces a model-agnostic Physics-Residual Learning framework that, when combined with few-shot calibration and diversity-oriented active sampling, significantly enhances the accuracy and stability of cross-geological Uniaxial Compressive Strength (UCS) predictions while reducing the need for destructive testing by over 90%.

Original authors: Long Long¹

Published 2026-10-21

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Long Long¹

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Deep underground, where tunnels are bored and foundations are laid, engineers face a fundamental question: how strong is the rock? The answer lies in a measurement called the uniaxial compressive strength, a number that tells us how much pressure a rock can withstand before it crumbles. This single value is the bedrock of safety calculations for massive infrastructure projects. Traditionally, finding this number requires a destructive process: drilling out a cylinder of rock, taking it to a laboratory, and crushing it in a machine. It is expensive, time-consuming, and often impossible if the rock is too fractured to be removed intact. For decades, scientists have tried to predict this strength using indirect clues, such as how fast sound waves travel through the rock or how dense it is. In recent years, powerful computer programs known as machine learning have been trained to spot patterns in these clues, promising to predict rock strength without ever breaking a sample.

However, a hidden flaw has plagued these digital predictions. Most computer models are trained on data from one specific location and then expected to work perfectly in a completely different geological setting. Just as a model trained on the weather in London might fail to predict a monsoon in Mumbai, a machine learning model trained on rocks in Brazil often stumbles when applied to rocks in Turkey. The geological differences are too vast, and the patterns the computer learned in one place do not translate to another. This creates a dangerous gap where engineers cannot rely on their predictions when they need them most: in new, unexplored terrain.

A new study by Long Long of the China Railway Corporation addresses this problem by changing how the computer learns. Instead of asking the machine to memorize the entire relationship between rock clues and strength from scratch, the researcher built a system that first applies a basic rule of physics. Think of this as giving the computer a rough map based on universal laws of nature, and then asking it to learn only the small, local corrections needed to make that map accurate for a specific country. The computer is no longer guessing the whole picture; it is simply adjusting a known baseline to fit the local reality. This approach, called physics-residual learning, separates the universal truth from the local quirks, making the model far more adaptable when it encounters a new geological landscape.

The researcher tested this idea using a massive collection of rock data from six different nations: Malaysia, Turkey, Iran, Portugal, South Africa, and Brazil. The goal was to see if a model trained on five of these countries could accurately predict the rock strength of the sixth, a scenario known as cross-geological prediction. The results revealed a surprising trade-off. When the model was asked to predict without any local data at all, the traditional computer-only approach actually performed slightly better. However, this initial advantage was an illusion. As soon as the model was allowed to look at just a few local samples to adjust its predictions, the physics-based approach surged ahead. The traditional model struggled to adapt, often becoming less accurate with just one or two new samples, while the physics-guided model used that small amount of data to correct its course with remarkable speed.

To measure exactly how well the models could adapt, the study introduced a new metric called the Calibratability Index. This number represents how much of the gap between the training data and the new location could be bridged using only five local samples. The physics-guided model managed to bridge over half of this gap, achieving a score of 0.511, whereas the traditional model only managed to bridge about 36 percent. This difference is critical because it means the new method requires far fewer expensive tests to become reliable. The study found that by using a smart selection process to choose which few samples to test, engineers could reduce the number of destructive rock-crushing tests needed by more than 90 percent. Instead of testing dozens of samples, a field team could select just three to five strategically chosen cores, run the tests, and let the model adjust the global prediction to fit the local site with high confidence.

The research also uncovered that not all rocks are created equal, and the errors in prediction are not random. The study calculated a "geological bias" for each country, revealing systematic differences in rock strength that the model could now interpret. For instance, the rocks in South Africa were found to be, on average, 5.6 times stronger than the global baseline, a reflection of their ancient, stable geological history. In contrast, the rocks in Turkey were found to be significantly weaker, only about a third of the baseline strength, likely due to their younger, more fractured nature. These findings confirm that the errors in prediction are not just noise; they are meaningful signals that reflect the true geological history of a region. By understanding these patterns, the model can be tuned to respect the unique character of each landscape.

One of the most practical discoveries was how to choose those few critical samples. The study compared picking samples at random against a method that looks for diversity in the rock's physical properties, such as density and wave speed, without needing to break the rock first. The diverse selection method proved far superior, improving prediction accuracy by 14 to 17 percent compared to random guessing. This means that engineers can use non-invasive tools to scan a site, identify the most representative spots, and then take just a handful of cores for testing. This strategy ensures that the small amount of destructive testing is used where it matters most, maximizing the value of every sample.

The study also tested whether adding more complex mathematical tricks or creating new features from the data would help. It turned out that simpler was better. The original set of measurements—wave speed, density, and rock type—performed just as well as, or better than, any complicated engineering of new variables. This suggests that for these types of geological problems, the most powerful tool is often the most direct one. The researchers found that trying to force the computer to learn more complex patterns with limited data actually led to worse results, confirming that a straightforward approach grounded in physical reality is more robust.

In the end, this work offers a clear path forward for geotechnical engineering. It demonstrates that by combining the universal rules of physics with the pattern-recognition power of machine learning, we can create models that travel well across different geological worlds. The method is not dependent on a single type of computer algorithm; it works with several different learning tools, making it a flexible framework for the industry. Most importantly, it provides a statistically proven way to drastically cut the cost and time of site investigation. By reducing the need for destructive testing from dozens of samples to just a few, this approach allows for faster, safer, and more economical engineering decisions in the complex and varied landscapes of the world.

Technical Summary: Physics-Residual Learning with Few-Shot Calibration for Cross-Geological UCS Prediction

Problem Statement
Predicting Uniaxial Compressive Strength (UCS) across geologically varied landscapes presents a significant challenge for data-centric machine learning (ML) models. While ML models often achieve high accuracy within a specific training domain, they frequently fail when extrapolated to unfamiliar geological contexts due to domain shift. Traditional approaches rely on random data partitioning, which assumes training and testing data share the same distribution, a condition rarely met in geotechnical engineering where models must be deployed in new locations with distinct lithological properties. Furthermore, existing physics-informed methods have not been systematically applied to address cross-geological extrapolation or the specific requirements of few-shot calibration, where destructive testing must be minimized.

Methodology
The study proposes a Model-Agnostic Framework centered on Physics-Residual Learning combined with Few-Shot Calibration. The methodology is structured as follows:

Dataset: The research utilizes the ROCK/10/4025 database, comprising 397 intact rock specimens from six nations (Malaysia, Turkey, Iran, Portugal, South Africa, Brazil) representing igneous, sedimentary, and metamorphic rock types. Input features include P-wave velocity ( $V_p$ ), dry density ( $\rho_d$ ), point load index ( $I_{s50}$ ), and rock classification.
Validation Strategy: A rigorous Leave-One-Country-Out (LOTO) protocol is employed. In each iteration, data from one country serves as the target domain (test set), while data from the remaining five countries form the source domain (training set). This simulates real-world cross-geological deployment.
Architectures: Four predictive frameworks are compared:
1. Direct ML: Standard data-driven models trained directly on raw features.
2. Physics-Guided Features (PGFE): Physics-based predictions are added as an input feature to the ML model.
3. Physics-Residual: The model predicts the residual (difference) between observed values and a physics-derived baseline, rather than the absolute UCS value.
4. Physics-Constrained Monotonic Residual: Similar to Physics-Residual but with monotonic constraints on physical features.
Physics Baseline: A tiered log-linear model based on physical principles (e.g., $log(\sigma_c) = a \cdot log(V_p) + b \cdot log(\rho_d) + c$ ) serves as the anchor. Coefficients are constrained to be positive to ensure physical plausibility.
Few-Shot Calibration: For each target domain, a small number of samples ( $k \in \{0, 1, 3, 5, 10, 20\}$ ) are selected from a calibration pool. A bias correction factor is calculated as the median difference between observed and predicted log-UCS values for these $k$ samples and applied to all predictions.
Active Sampling: Two strategies are compared for selecting the $k$ calibration samples: Random Sampling and Diversity Sampling (using $k$ -means clustering on non-invasive features $V_p$ and $\rho_d$ to ensure feature space coverage).
Metrics: Performance is evaluated using RMSE, MAE, $R^2$ , and a novel Calibratability Index (CI), which quantifies the proportion of the domain gap bridged by calibration.

Key Contributions

Physics-Residual Framework: The study demonstrates that anchoring ML models to physical laws transforms unpredictable extrapolation errors into systematic, interpretable biases that can be corrected with minimal data.
Calibratability Index (CI): A new metric is introduced to measure the efficiency of domain adaptation. The Physics-Residual approach achieves a CI of 0.511, significantly outperforming Direct ML (0.364), indicating a 40% improvement in calibration efficiency.
Diversity-Oriented Active Sampling: The research shows that selecting calibration samples based on the diversity of non-invasive measurements ( $V_p$ , $\rho_d$ ) reduces RMSE by 14–17% compared to random sampling, allowing for reliable predictions with only 3–5 destructive tests.
Geological Bias Correction (GBC): The study quantifies systematic cross-domain errors, revealing that bias multipliers vary geologically (e.g., 0.36 for Turkey vs. 5.60 for South Africa), confirming that cross-domain errors are structured and geologically significant rather than random noise.
Parsimony Validation: Feature engineering (e.g., adding dynamic modulus or log-transformations) was found to offer no improvement over the original feature set, reinforcing the principle of parsimony for small geotechnical datasets.

Results

Zero-Shot vs. Calibrated Performance: In zero-shot scenarios ( $k=0$ ), Direct ML initially showed a slightly lower RMSE (58.5 MPa) compared to Physics-Residual (60.2 MPa). However, this advantage vanished immediately upon calibration.
Calibration Efficiency: At $k=3$ , Physics-Residual achieved an RMSE of 37.7 MPa, outperforming Direct ML (41.1 MPa). At $k=5$ , the RMSE dropped to 36.8 MPa with Diversity sampling, representing a 17% improvement over random sampling.
Stability: A "k=1 instability" phenomenon was observed where single-sample calibration increased error for all models, suggesting a minimum budget of $k \ge 3$ is required for reliable correction.
Statistical Significance: Wilcoxon signed-rank tests confirmed the superiority of Physics-Residual over Direct ML with $p < 10^{-8}$ and medium effect sizes ( $d \approx -0.53$ to $-0.58$).
Geological Insights: The GBC multipliers aligned with geological realities; South African rocks (cratonic, low porosity) required a 5.6x upward correction, while Turkish rocks (young orogenic belt) required a 0.36x downward correction.

Significance and Claims
The paper claims to provide a feasible, statistically robust process for field implementation that reduces the need for destructive UCS testing by over 90% (from 30–50 tests to just 3–5). By converting erratic extrapolation errors into systematic biases, the framework enables dependable UCS forecasting in new geological contexts without requiring extensive local data collection. The author emphasizes that the methodology is model-agnostic, offering consistent benefits across various tree-based learners (CatBoost, XGBoost, Random Forest, LightGBM). The introduction of the Calibratability Index is presented as a universal tool for assessing domain adaptation efficiency in any physical science field requiring localized empirical adjustments. The study concludes that while feature engineering is often pursued, the integration of physical principles and strategic sampling is the primary driver of success in cross-geological prediction.

Technical Summary: Physics-Residual Learning with Few-Shot Calibration for Cross-Geological UCS Prediction

More like this