Kerogen is the primary hydrocarbon source of shale oil/gas. The kerogen types and maturity are the two most crucial indicators that can reflect the hydrocarbon generation potential of shale oil/gas reservoirs. These indicators and the other mechanochemical properties can be effectively studied in a bottom-up strategy using kerogen molecular models. Thus, the rapid construction of kerogen molecular models is the cornerstone of shale oil/gas exploitation research. Because of the combinatorial explosion problem, there are two inherent disadvantages of traditional methods: being time- and material-consuming and labor-intensive. We propose a new method that combines machine learning with multiple nuclear magnetic resonance spectra to intelligently and with a high throughput predict the kerogen structures, types, and maturity. Neither the manual analysis of experimental spectra nor the enormous trial-and-error process is required in our method. The 650,000 groups of samples are annotated as the sample datasets. Various spectral types can be analyzed comprehensively using the multi-spectral form, and the predictive capability beyond that of the single input form is obtained. The results demonstrate that the average similarity of prediction molecules and the targets is 91.78%. The prediction accuracy of kerogen components, types, and maturity indexes is better than 92.4%, and the coefficients of determination R2 are all over 0.934. The results exhibit the excellent comprehensive performance and effectiveness of our method. Thus, we anticipate that this work will shorten the research cycle and tremendously reduce costs in constructing kerogen models and predicting kerogen properties.
Read full abstract