Predicting Radon Emissions from Underground Coal Fires Using Machine Learning

Beneath the scorched earth of an abandoned open-pit mine in northeastern China, an invisible tracer is leaking toward the surface, and a team of researchers has now taught an algorithm to read it. Radon, the radioactive noble gas produced by the decay of uranium in rock, has long been recognized as a sentinel of underground coal fires: as combustion fronts carve fracture networks through buried coal seams, they open highways for radon to migrate upward, producing measurable anomalies in surface air. The problem has always been that radon measurements are slow, expensive, and heavily confounded by weather. A new study published in Natural Resources Research by Hemeng Zhang of Liaoning Technical University and colleagues demonstrates that machine learning can cut through that atmospheric noise, achieving a coefficient of determination of 0.981 on unseen test data and, in the process, offering a cheaper, faster route to locating and monitoring one of the world’s most persistent industrial hazards.

Underground coal fires are a quietly catastrophic phenomenon. Across coal-producing regions from China to India, the United States, and South Africa, seams that ignited spontaneously or through human activity can smolder for decades, consuming irreplaceable energy resources, releasing enormous quantities of carbon dioxide and toxic gases, destabilizing the ground above, and rendering land unusable. Because the combustion zone is hidden from view, accurate prediction of the fire’s state—where it is burning, how intensely, and in which direction it is spreading—is critical for assessing hazard levels and designing suppression campaigns. Traditional detection methods range from thermal infrared remote sensing to magnetic and electrical geophysical surveys, and in situ gas measurements, but each carries limitations of cost, resolution, or depth sensitivity.

Radon has emerged as one of the most effective chemical indicators precisely because of its physical behavior. The gas is chemically inert, so it is not consumed or transformed as it travels; it is generated continuously in rock and soil at depths of meters to hundreds of meters; and its half-life of about 3.8 days is long enough to survive migration from depth yet short enough that a strong surface signal implies an active, nearby source. Heat from a coal fire drives convection of pore air and dries and fractures the overlying strata, enhancing both radon production and its upward transport. Earlier studies, including work by the same research group on fracture-driven radon anomalies in coal fire zones, established that surface radon concentrations above burning goafs—abandoned mined-out areas—can be several times higher than background. But every measurement campaign has had to grapple with a fundamental complication: the atmosphere itself.

Surface radon readings are modulated by ambient temperature, barometric pressure, relative humidity, and wind. Temperature swings alter thermal gradients and convection within the soil column; pressure changes pump soil gas alternately inward and outward; humidity affects the adsorption and diffusion of radon through pore water; and wind dilutes the near-surface gas layer. Because these meteorological variables oscillate on diurnal and weather-front timescales, a single radon reading is a tangled superposition of the geological signal of interest and environmental interference. Field crews must therefore repeat measurements under multiple conditions or laboriously correct for weather, which is exactly what makes radon surveying time-consuming and costly at the scale needed for operational fire monitoring.

The team’s response was to invert the problem. Rather than trying to remove the atmospheric signal, they taught statistical models to predict it. The researchers collected 216 paired samples from the fire area of the Haizhou open-pit mine, a well-characterized site where underground combustion has been documented, recording surface radon concentration alongside four meteorological variables: ambient temperature, ambient pressure, relative humidity, and wind velocity. With this dataset they trained and compared four machine learning algorithms spanning a spectrum of complexity. At the simple end sat ridge regression, a regularized linear method that penalizes large coefficients to combat multicollinearity among the atmospheric predictors, following the classic formulation of Hoerl and Kennard. At the complex end stood three tree-based ensemble methods: random forest, which aggregates many decision trees trained on bootstrap resamples of the data; extra trees, an even more randomized variant that splits nodes on randomly drawn thresholds; and gradient boosting, which builds trees sequentially, each new learner fitted to the residual errors of the ensemble so far, in the spirit of Friedman’s gradient boosting machine.

The evaluation protocol compared the models on held-out testing data using three complementary metrics. Gradient boosting emerged as the clear winner, achieving an R² of 0.981, a root-mean-square error of 170.29, and a mean absolute error of 124.69 on the test set—an exceptional fit for field measurements of a naturally variable gas. The results demonstrated that the gradient boosting model exhibited exceptional fit and robust generalizability, maintaining high stability and accuracy when asked to predict radon concentrations over different time durations, which the authors interpret as evidence that the model captured genuine atmosphere–radon coupling physics rather than memorizing the specific campaign’s noise. The extra trees and random forest ensembles performed respectably, while ridge regression, as expected for a linear model confronting nonlinear interactions between temperature, humidity, pressure, and soil-gas transport, lagged behind.

Perhaps the most operationally significant finding came from the sensitivity analysis: relative humidity was the single most influential factor governing surface radon anomalies in the fire zone. This makes physical sense. Moisture in the pore space of the overburden strongly governs whether radon atoms, born in mineral grains, are able to escape into pore air and migrate upward, and humidity also modulates the thermal contrast between the burning seam and the surface that drives convective pumping. In practical terms, the result means that radon-based fire surveys should be interpreted through the lens of humidity conditions at the time of measurement—and, better still, that humidity data, which are cheaply available from standard weather stations, can be combined with a trained gradient boosting model to correct or even substitute for some radon measurements.

The implications for coal fire management are substantial. If surface radon anomalies can be predicted from routinely monitored atmospheric variables, then the inverse problem—inferring the state of the underground fire from deviations between predicted and observed radon—becomes tractable. A persistent gap between the model’s expectation and the actual reading could flag a newly igniting fire zone, a change in the combustion front’s intensity, or the development of new fracture pathways, all without waiting for slow, laboratory-based radon assay. Because the framework requires only four standard meteorological inputs, it could be deployed alongside existing weather stations at coalfields in China, India, Indonesia, South Africa, and elsewhere, providing continuous, near-real-time hazard assessment at a fraction of the cost of dense manual radon grids.

The study also slots into a broader wave of machine learning applications in geohazard and subsurface monitoring that the authors and their collaborators have helped drive. The same group has previously applied optimized ensemble models to predict carbon dioxide fluxes from underground coal fires in support of net-zero carbon accounting, used advanced regression schemes for shale wettability prediction in carbon capture and storage contexts, and explored artificial intelligence forecasting for methane recovery from gas hydrates and CO2 storage efficiency in saline aquifers. The present work extends that toolkit to a public-safety and environmental-protection application, and the authors suggest it opens new directions for applying machine learning in environmental fields more broadly, from soil gas monitoring to volcanic and geothermal degassing studies where radon and CO2 serve as structural tracers.

Limitations remain, and the authors are measured about them. The model was trained at a single site, the Haizhou fire area, and while its performance held across different time windows there, transfer to other geological settings—different overburden lithologies, fire depths, and climatic regimes—will require site-specific retraining, since the relationship between humidity, pressure, and radon exhalation depends on local pore structure and fracture geometry. The dataset of 216 samples, while adequate for four predictors, is modest by deep learning standards, which is one reason the researchers favored gradient boosting: the method delivers strong tabular-data performance with relatively few samples and provides interpretable sensitivity rankings. Future work, the team indicates, will aim at integrating the radon prediction framework with other geophysical indicators and expanding validation across multiple coalfields.

For now, the study stands as a demonstration that the weather itself, so often the enemy of gas-based fire detection, can be turned into an ally. By learning the atmospheric fingerprint superimposed on the radon signal, a gradient boosting model trained at an old Chinese open-pit mine turned four ordinary weather variables into a window on a hidden fire, with accuracy above 98 percent. If the approach generalizes, the slow, expensive business of chasing radon anomalies across burning coalfields may soon be replaced by something far more agile: an algorithm, a weather station, and a sense of where the ground below is quietly burning.

Zhang, H., Zhang, Z., Vo Thanh, H., Jia, S., Dai, Z., Zhang, Y., & Wang, Y. (2026). Machine learning-based prediction of radon emissions from underground coal fires based on atmospheric changes. Natural Resources Research. https://doi.org/10.1007/s11053-026-10748-2