Case Study: Using AI&ML to Generate Well Logs in Santa-Fe Field, Kansas

Well logs are one of the most crucial field measurements in formation evaluation and reservoir characterization. They have been utilized for almost a century since 1926 when Schlumberger brothers invented the electric logs. During the decades, well logging tools and interpretation techniques have gone under many advancements and developments adding to its value as a critical field measurement. However, there are several thousand cases in which due to issues such as technical problems of the tools, operational state of the well, age of the well, time sensitivity, and economical reasons few or many of the logs are not measured. In addition to missing logs, there are many cases that logs are measured but have a bad quality and are not reliable. Nowadays, Artificial Intelligence and Machine Learning (AI&ML) techniques can be used to provide the missing data and/or identify bad quality ones. There are several studies in the literature that have used multiple available logs and trained different AI&ML models like Artificial Neural networks (ANN), Decision Tree, Random Forests, XGBoost to predict a single missed well log i.e., Sonic, Neutron Porosity, Density. In other words, multiple available logs were used as input to predict a single output log for a well. However, in reality thousands of cases can be found where a well misses multiple logs or have the minimum available logs. This study will address this issue. In this case study which is used on Santa Fe field data located in Kansas, a sequence of nine ANNs were trained, Calibrated, validated, and tested (blind validation) to generate synthetic logs step-by-step only based on GR (Gamma Ray) and caliper logs. This study consists of ten wells where nine of them were used for training, calibration, validation, and one well has been used for blind validation of the ANNs performance. The blind validation well is a well that was never used in any of ANNs training, calibration, and validation and is assumed not to exist. The order of nine ANNs sequence is to predict Neutron Porosity, Sonic Travel Time, Density, Photoelectric, and Array Resistivities (R10, R20, R30, R60, R90), respectively. The order of ANNs sequence is determined based on logs' "Key Performance Index (KPI)", a technique based on Fuzzy Set Theory. In all nine ANNs prediction of the blind validation well, only GR and Caliper measurements are used, and any other logs that are used as an input are generated synthetically by the previous ANNs. The results successfully show that all missed logs can be generated with high accuracy based on minimum available logs like GR and Caliper.