.png){kind=icon}
.png){kind=icon}
Return
Optimizing feature selection for global geothermal heat flow prediction using machine learning
.png){kind=icon}
OA
AI
DOI：10.1038/s41598-026-63610-z.png)
.png){kind=icon}
Original request for help.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Accurate prediction of surface geothermal heat flow (GHF), which reflects Earth’s internal heat loss, is essential for assessing geothermal energy potential and supporting low-carbon energy development. Machine learning is increasingly applied in regions lacking direct measurements; however, feature selection is often arbitrary or not geothermally grounded, and the perceived need for extensive input data can hinder model application, while the use of poorly constrained variables may increase prediction error. This study presents a systematic workflow to identify consistently important predictors for GHF by integrating feature importance, correlation analysis, permutation importance, feature-wise evaluation metrics, recursive feature elimination, column-drop analysis, and automated selection methods, all assessed using cross-validation, out-of-bag error, and early stopping. Our results indicate that reliance on individual feature-ranking methods can be misleading and often results in biased interpretations. Using an integrated feature selection framework, six dominant predictors for GHF consistently emerge: distance to 5 nearest volcanoes, hotspots, and trenches, along with heat production provinces, Moho depth, and lithosphere–asthenosphere boundary (LAB) depth, all of which are physically linked to heat flow processes. Using this reduced feature set, we generate a new global GHF prediction for continental regions with uncertainty of ca. 8 mW/m2, from the ensemble prediction variability of the Gradient Boosting Regressor. The proposed workflow provides a transferable framework for machine learning-based prediction across scientific disciplines.
Journal
IF:
3.9
Papers:
27.4W
Citations:
83.5W