📄 earth_science

Permeability Evolution and Cross-Specimen Machine-Learning Prediction of Geothermal-Reservoir Sandstone under Coupled Temperature and Pressure

This study develops a robust machine-learning framework using nested specimen-grouped cross-validation to predict the path-dependent permeability of geothermal-reservoir sandstone under coupled temperature and pressure, identifying random forest as the optimal model and highlighting confining pressure as the most stable predictive factor while demonstrating incomplete permeability recovery after unloading.

Original authors: Jibin He¹˒², Li Yu¹, Kun Zhao¹, Xiaoyan Niu¹

Published 2026-08-10

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Jibin He¹˒², Li Yu¹, Kun Zhao¹, Xiaoyan Niu¹

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

Imagine the Earth's crust as a giant, ancient sponge, but instead of soaking up water from a rainstorm, it holds onto scorching hot water deep underground. This is the world of geothermal energy, where we try to tap into the planet's natural heat to power our homes. To do this, engineers drill deep wells and pump cold water down, hoping it will heat up as it squeezes through tiny cracks and holes in the rock, then rise back up as steam or hot water. But here's the tricky part: the rock isn't just a static sponge. It's a living, breathing structure that changes shape when you heat it up, squeeze it with heavy pressure, or push water through it. If the rock gets too squished, the tiny tunnels (called pores) can close up, blocking the flow. If it gets too hot, the minerals might expand or crack, opening new paths or clogging old ones. Understanding how these underground rocks react to this mix of heat and pressure is like trying to predict how a rubber band will stretch if you heat it while pulling it. If we get this wrong, the energy wells could clog up and stop working, wasting millions of dollars and leaving us without clean power.

This is exactly the puzzle a team of researchers from Hebei University set out to solve. They weren't just guessing; they were playing a high-stakes game of "guess the rock's mood" using a special kind of computer brain called machine learning. They took 16 real chunks of sandstone from a geothermal site in Jinan, China, and put them through a rigorous workout. They heated them up to temperatures as high as 80°C, squeezed them with pressures up to 19 MPa, and pumped water through them at different speeds. They did this over and over, loading and unloading the pressure like a stress test, collecting 336 different measurements of how easily water could flow through the rock.

The big challenge they faced was a sneaky trap that often tricks computer models: "data leakage." Imagine you are trying to predict how a specific student will do on a math test. If you let the computer study that student's homework and then test it on the same student's quiz, the computer will get a perfect score, but only because it memorized the answers, not because it learned math. In rock science, if you mix measurements from the same rock sample into both the "learning" and "testing" groups, the computer just learns the unique personality of that one rock, not how rocks in general behave. The researchers realized that to make a model that could predict how new, unseen rocks would behave, they had to be strict. They treated each rock specimen as a separate "student" and made sure the computer never saw the test rock's homework during its training.

Using this strict "no mixing" rule, they trained five different types of machine learning models to predict the rock's permeability (how easily water flows through it) based on four inputs: temperature, injection pressure, squeezing pressure, and how porous the rock was. The winner was a model called "Random Forest." Think of this model not as a single genius, but as a crowd of 100 different experts. Each expert looks at the data a little differently, and they vote on the answer. This "crowd wisdom" approach turned out to be the best at guessing how a brand-new rock would react, achieving a success rate (R²) of 0.748. This means the model could explain about 75% of the changes in the rock's behavior, which is a huge improvement over older, simpler methods that assumed the relationship was just a straight line.

One of the most interesting discoveries was about the "memory" of the rock. When they squeezed the rock and then let go, the rock didn't snap back to its original state perfectly. It was like stretching a piece of chewed gum; when you let go, it stays a bit stretched. They found that after unloading the pressure, the rock's ability to let water through was about 4.41% lower than when they were loading it. This "incomplete recovery" happened mostly during the first squeeze, suggesting that the first time you stress the rock, it rearranges its internal cracks and pores permanently.

The researchers also built a user-friendly tool, like a digital dashboard, where engineers can plug in numbers for temperature and pressure to get a quick estimate of how the rock will behave. However, they are very careful to say this tool only works for rocks that look and act like the ones they tested (sandstone from that specific area in Jinan) and within the specific temperature and pressure ranges they used (25–80°C and 10–19 MPa). It's not a magic crystal ball for every rock on Earth, but it is a powerful, honest tool for understanding the specific rocks they studied. By proving that you can't just mix up data from the same rock, and by showing that a "crowd" of computer models can predict rock behavior better than a single formula, this study gives engineers a more reliable way to design geothermal systems that won't get clogged up by the very rocks they are trying to harness.

Technical Summary: Permeability Evolution and Cross-Specimen Machine-Learning Prediction of Geothermal-Reservoir Sandstone

Problem Statement
Deep geothermal reservoirs rely on sustainable fluid injection and production, processes that subject reservoir sandstone to coupled thermal, hydraulic, and mechanical (THM) disturbances. These conditions alter pore-fracture structures, leading to path-dependent permeability changes that are difficult to capture using simple empirical relationships or single-variable functions. A critical methodological challenge in applying machine learning (ML) to this domain is data leakage: conventional random row-wise splitting of datasets often allocates multiple sequential measurements from the same rock specimen into both training and test sets. This allows models to "memorize" specimen-specific characteristics rather than learning generalizable physical laws, resulting in overoptimistic performance estimates that fail when predicting truly unseen specimens.

Methodology
This study addressed these challenges by developing a machine-learning framework based on 336 coupled temperature–pressure seepage measurements from 16 sandstone specimens sourced from a geothermal reservoir in Jinan, China. The experimental design covered temperatures of 25–80°C, injection pressures of 3–9 MPa, and cyclic confining pressures of 10–19 MPa, with 21 sequential measurements recorded per specimen.

The core methodological innovation was the implementation of nested specimen-grouped cross-validation to prevent data leakage:

Outer Loop: An eight-fold GroupKFold procedure held out complete specimens (two per fold) to evaluate the model's ability to predict previously unseen rocks.
Inner Loop: A five-fold grouped cross-validation on the remaining specimens optimized hyperparameters without mixing observations from the same specimen across folds.

Five regression models were evaluated: Multiple Linear Regression (MLR), Radial-Basis-Function Support Vector Regression (SVR), Random Forest (RF), Extremely Randomized Trees (Extra Trees), and Gradient Boosting Regression (GBR). Inputs included temperature, injection pressure, confining pressure, and porosity; the output was permeability. Model interpretability was assessed via permutation importance on held-out specimens and SHapley Additive exPlanations (SHAP) analysis. Additionally, the study analyzed cyclic permeability recovery by comparing loading and unloading paths at matched confining pressures.

Key Results

Model Performance: Under strict specimen-grouped validation, tree-ensemble models significantly outperformed linear and kernel-based baselines. Random Forest (RF) achieved the best overall performance with a pooled out-of-fold $R^2$ of 0.748, Mean Absolute Error (MAE) of 0.0144 mD, Root Mean Square Error (RMSE) of 0.0190 mD, and Mean Absolute Percentage Error (MAPE) of 15.29%. Gradient Boosting and Extremely Randomized Trees produced comparable results, while MLR and SVR failed to generalize effectively across specimens.
Predictor Importance: Permutation analysis identified confining pressure as the most stable predictive source for unseen specimens. The importance of temperature, injection pressure, and porosity was found to be less stable, influenced by specimen differences and confounding among factor combinations. SHAP analysis confirmed that increasing confining pressure suppresses permeability, while high injection pressure generally increases it; however, these represent statistical associations rather than established physical causality.
Cyclic Hysteresis: Analysis of 144 loading–unloading pairs revealed incomplete permeability recovery. The median permeability difference (unloading minus loading) was −0.0040 mD (−4.41%). This irreversible loss was most pronounced during the first cycle (median difference ~−8%) and attenuated in subsequent cycles, indicating that initial loading causes significant fracture closure and grain rearrangement, after which the internal structure stabilizes.
Residual Heterogeneity: Prediction errors were specimen-dependent rather than purely random, with specific specimens showing systematic over- or under-prediction, highlighting the influence of intrinsic pore structure and initial fracture states.

Significance and Claims
The paper claims its primary contribution is not merely a higher coefficient of determination, but a redefinition of the model-validation target for repeated measurements in rock physics. By shifting the evaluation from interpolation within known specimens to transfer prediction across unseen specimens, the study provides a more credible assessment of model generalization.

The authors emphasize that using specimens rather than individual measurements as validation units effectively limits data leakage. The study concludes that tree-ensemble models are better suited for datasets with a small number of independent specimens and repeated measurements. Furthermore, the development of a permeability prediction interface based on the final RF model offers a practical tool for rapid permeability assessment within the specific experimental domain (25–80°C, 3–9 MPa injection, 10–19 MPa confining, 17.2–19.2% porosity). The authors modestly note that this tool is an auxiliary decision support system restricted to the specific geothermal sandstone batch studied and does not serve as a universal predictor for all rock types or conditions.

Technical Summary: Permeability Evolution and Cross-Specimen Machine-Learning Prediction of Geothermal-Reservoir Sandstone

More like this