The exploitation of hydrocarbon reservoirs may potentially lead to contamination of soils, shallow water resources, and greenhouse gas emissions. Fluids such as methane or CO2 may in some cases migrate toward the groundwater zone and atmosphere through and along imperfectly sealed hydrocarbon wells. Field tests in hydrocarbon-producing regions are routinely conducted for detecting serious leakage to prevent environmental pollution. The challenge is that testing is costly, time-consuming, and sometimes labor-intensive. In this study, machine learning approaches were applied to predict serious leakage with uncertainty quantification for wells that have not been field tested in Alberta, Canada. An improved imputation technique was developed by Cholesky factorization of the covariance matrix between features, where missing data are imputed via conditioning of available values. The uncertainty in imputed values was quantified and incorporated into the final prediction to improve decision-making. Next, a wide range of predictive algorithms and various performance metrics were considered to achieve the most reliable classifier. However, a highly skewed distribution of field tests toward the negative class (nonserious leakage) forces predictive models to unrealistically underestimate the minority class (serious leakage). To address this issue, a combination of oversampling, undersampling, and ensemble learning was applied. By investigating all the models on never-before-seen data, an optimum classifier with minimal false negative prediction was determined. The developed methodology can be applied to identify the wells with the highest likelihood for serious fluid leakage within producing fields. This information is of key importance for optimizing field test operations to achieve economic and environmental benefits.

We demonstrate the usefulness of cognitive models for combining human estimates of probabilities in two experiments. The first experiment involves people’s estimates of probabilities for general knowledge questions such as “What percentage of the world’s population speaks English as a first language?” The second experiment involves people’s estimates of probabilities in football (soccer) games, such as “What is the probability a team leading 1–0 at half time will win the game?”, with ground truths based on analysis of large corpus of games played in the past decade. In both experiments, we collect people’s probability estimates, and develop a cognitive model of the estimation process, including assumptions about the calibration of probabilities and individual differences. We show that the cognitive model approach outperforms standard statistical aggregation methods like the mean and the median for both experiments and, unlike most previous related work, is able to make good predictions in a fully unsupervised setting. We also show that the parameters inferred as part of the cognitive modeling, involving calibration and expertise, provide useful measures of the cognitive characteristics of individuals. We argue that the cognitive approach has the advantage of aggregating over latent human knowledge rather than observed estimates, and emphasize that it can be applied in predictive settings where answers are not yet available.