ABSTRACT

In salt basins, anticipating the spatial repartition of intra-salt lithologies is a prerequisite for safe drilling operations and for the safe development of salt caverns. However, it requires to compile a large set of well data in the area of interest and because of bad cuttings recovery during drilling operation, the identification of lithologies is a cumbersome task done manually combining well events, logs and regional context. One challenge is thus to automate and standardize evaporites identification from log data.

Automated lithology identification from logs is a problem generally addressed for sedimentary lithologies (e.g., claystone, sandstone etc.); it requires a sound understanding of wellbore context and log measurements. Recent applications with Machine Learning (ML) have shown that either classification or regression methods can deliver accurate results in some specific contexts. But predictions can rapidly degrade in case of severe wellbore artifacts and for lithologies poorly represented in the training dataset. Evaporites combine generally the two issues simultaneously with possibly poor borehole conditions and a large variety of lithologies seldomly distributed in log database. Accordingly, ML Classification with a high accuracy in evaporites is challenging and innovative.

A database was built with two famous areas for evaporites: North Sea and Brazil. It includes a total of 743 wells having expert-validated evaporite lithology flags and a uniform set of conventional logs: Density, Neutron, Gamma Ray and Compressional Velocity logs. Log signatures were treated based on wellbore quality; data were filtered by thickness, distance to boundaries and wellbore quality (diameter and rugosity). Then, considering that evaporites are pure nonporous lithologies, the Euclidian multidimensional distance and dispersion of the log data to the tabulated evaporite log coordinates were taken as a probabilistic quality indicator. Finally, due to some possible lithology mixtures (mainly Halite and Anhydrite), a few lithologies were reallocated, subdivided (Sylvite split per GR and acoustic signature) or grouped (Carnallite and Bischofite). The final database contains eight salt lithology names (Potassic salt, Anhydrite, Halite, Sylvite, Carnallite, Polyhalite, Bischofite and Tachyhydrite) covered by same logs and having equivalent clustering ratio. This represents approximately 10% of the initial dataset for a total of 400,000 training points.

A large panel of ML clustering approaches were assessed with AutoML to select the best performing and most robust model (Random Forest). The optimal combination of input logs was also tested and came, by order of importance, to Density, Gamma Ray, Neutron, Compressional and shear velocities. In case a log was missing, a synthetic ML-derived version was used. After this exploration step, predictions were further improved with an average-standard deviation rescaling for each log and grid search for hyperparameters optimization. Since lithologies were quite imbalanced in the original dataset, a class weight, inversely proportional to the cardinality of the target class, was applied for each depth frame. The model validation was performed with a set composed of entire wells having evaporite lithologies in same frequency as the original raw dataset. Finally, some postprocessing was performed on outputs to reallocate some classes which increased the prediction accuracy.

The full workflow was applied to the North Sea and Brazilian sectors. and the model could predict with a high accuracy:

• Anhydrite – Halite > 90%

• Sylvite – Carnallite > 70%

The results were used to define intra-salt lithology distribution, characterize evaporites acoustic properties and initialize seismic models. The whole study cycle could be achieved thanks to the automation brought by the ML steps, which, in return, allowed an exhaustive use of the well data and ensured interpretation consistency all over the area.

This content is only available via PDF.
You can access this article if you purchase or spend a download.