Research progress and application prospect of interpretable machine learning in artificial intelligence of oil and gas industry

Artificial intelligence, as a strategic emerging industry and a new quality productivity, is rapidly penetrating into the oil and gas industry, and is expected to become a new engine and commanding elevation for the development of the industry. The machine learning model of black box is lack of transparency and interpretability, which leads to low acceptance and trust of existing machine learning methods in the oil and gas industry and restricts the fusion and development of the artificial learning with machine learning as the core in oil and gas fields. In this paper, the research status of interpretable machine learning method in the process of oil and gas field exploration and development is introduced systematically. It is indicated that the interpretability of machine learning model is the key to promote the large-scale application of artificial intelligence in the oil and gas industry. Then, the limitations of the post-hoc interpretability method in oil and gas machine learning method are illustrated, and the application of the technology is predicted. And the following research results are obtained. First, the post-hoc interpretability methods such as Shapley Additive Explanation (SHAP) and Local Interpretable Model-agnostic Explanation (LIME) are used for the case verification of the main control factors of CBM productivity. It is suggested that the interpretable oil and gas field characteristic indexes are inadequate to fully guide the construction and analysis of interpretable model. Therefore, an intrinsic interpretability machine learning method which is accordant with the characteristics of oil and gas field exploration and development itself shall be established based on the idea of intrinsic interpretability. Second, the intrinsic interpretability methods such as mechanism model, causal inference, and counterfactual explanation are used to analyze the causality between oil and gas field data and model parameters, and then the intrinsic interpretability machine learning method is constructed. Third, the typical CBM fracturing data is selected for the case verification of productivity prediction. It is indicated that causal inference can effectively mine the intrinsic relationships between geological parameters, construction parameters and production capacity, and the machine learning model established based on causality can improve the generalization of prediction. In conclusion, the machine learning method based on post-hoc interpretability and intrinsic interpretability is not only the inevitable development trend of artificial intelligence in the oil and gas industry in the future, but also the bottlenecks and key technology for the field application of artificial intelligence in the oil and gas industry. © 2024 Natural Gas Industry Journal Agency. All rights reserved.