Development of machine learning models to determine CO2 seepage in the ocean based on the seepage and non-seepage database.

CCS is gaining attention as a measure against the global warming. In Japan, a CCS operator must conduct monitoring to check whether CO2 stored under the seabed safely and, if high CO2 concentration in seawater is detected, reinvestigation is legally required. In the CCS demonstration project conducted offshore Tomakomai, Japan, seawater monitoring continues even after the injection well was closed in 2019. Since the project began, false positives have frequently led to the need for reinvestigation. This is because the indicator used is based on the criterion whether data points exceed the upper limit of the 95% prediction interval of data variation. There is no scientific basis for the 95% figure. The objective of this study is to construct machine learning models which can detect CO2 seepage with higher accuracy. To do so, we conducted water sampling from seas where CO2 seeps artificially and naturally. However, since the number of seepage data was limited: namely, 29 seepage data and 824 non-seepage data, we adopted the synthetic minority over-sampling technique and increased the number of the minority class to 169. As a result, the Random Forest and the Feedforward Neural Network model achieved F-scores of 0.982 and 0.981, respectively, exceeding the score of 0.925 achieved by the current standards of the Ministry of Environment, Japan. By using these models and the database, it is expected that CO2 seepage in the ocean can be determined with high accuracy.