Work

Uncertainty quantification of geochemical data imputation using Monte Carlo dropout

Year: 2025

Type: article

Abstract: Machine learning models have shown their promise in geochemical data imputation tasks. However, being black-box solvers, these models require more confidence in their predictions. Using uncertainty quantification methods for deep neural networks can increase the reliability of their predictions. In this paper, we use Monte Carlo Dropout to estimate uncertainty in geochemical data imputation. Multiple forward passes with different dropout configurations yield a predictive distribution for the unknown analytes. The mean of this distribution is used as the prediction, while the standard deviation expresses the uncertainty of the neural networks. Two different scenarios, namely the WACHEM and WAMEX databases containing multi-element geochemical data for rock samples, illustrate the predictive accuracy of the method and its capability to measure the associated uncertainty. Dropout values of 0.1–0.2 were identified as a good balance in prediction accuracy and model uncertainty. • Monte Carlo Dropout is used to estimate uncertainty in geochemical data imputation. • The method has low uncertainty for samples with high information content and vice versa. • Dropout values of 0.1–0.2 are a good balance in prediction accuracy and model uncertainty.

Source: Journal of Geochemical Exploration

DOI: 10.1016/j.gexplo.2025.107695

Authors: Vladimir Puzyrev, Paul Duuring

Institutions: Curtin University, Geological Survey of Western Australia, The University of Western Australia, Department of Local Government, Industry Regulation and Safety

Language: English

Topic: Geochemistry and Geologic Mapping

Subfield: Artificial Intelligence

Field: Computer Science

Domain: Physical Sciences

Open Access status: hybrid

Estimated APC paid: $3,570

Funders: U.S. Department of Energy, Department of Mines, Industry Regulation and Safety, Government of Western Australia, Geological Survey of Western Australia