The workflow and software tools used in this project were designed to handle large data sets (hundreds of wells) consistently and efficiently, with minimal guidance or intervention from the user. The output from the system includes quality assessments based on analysis of the input data and ML answers, with flags that identify those wells that need to be reviewed by the petrophysicist. The user has the option to process each well in its entirety (the holistic mode) or switch to a zone-by-zone process, provided carefully selected wells tops (or markers) are available. Typical LQC logs, such as Caliper, cable tension, density-log correction, are not required for executing the workflow, although they can be used if and when available.

The workflow involves five basic steps, illustrated in the Fig.2 below. Further details of the algorithms, with examples, can be found in Akkurt et al. (2023). 

The first step of the workflow involves loading and scanning of the input data. A series of basic checks (consistency of sampling rates, identification of dead traces, etc.) are carried out in this step to ensure that data is properly harmonized. Normalization of certain variables, as well as creation of additional features (differential hole size, resistivity ratios, various flags for QC) are also completed in this step. The user can quickly get an idea of the overall data availability and quality at the end of this step. An example of log availability for a subset of the data can be seen in Fig.3. Outliers, defined as the data points that are inconsistent with the rest of the data, are identified in Step 2 using a combination of concepts borrowed   from the field of Outlier Detection, operating in multi-dimensional space. A discussion of the outlier detection algorithm can be found in Akkurt et al (2022). In Step 3, data free of outliers are used to define the petrophysical footprint of each well, which is used to quantify petrophysical similarity between wells. Various similarity metrics are then consumed to create clusters of wells based on similarity. This step is designed to handle spatial variations and can be considered an ‘analog-well’ finder. 

Step 4 involves training and regression, where for each well, ML models are trained to a cluster of wells as determined in Step 3 and applied to predict the required log responses. Step 5, the final step in the workflow, is crucial as it allows the user to assess the quality of the reconstruction carried out in Step 4. Automated analysis of uncertainty, as well as other QC metrics provided, are used to flag wells for further attention of the user. Furthermore, the wells are ranked, going from higher to lower uncertainty, to aid the user in the visual inspection process.