100 wells were selected for the project, located throughout the field without reference to a specific block. The data was limited to zones of interest – Horizons 12–18. The standard logging suite included: spontaneous potential (SP), gamma-ray (GR), neutron logs (TNPH), bulk density (RHOB), lateralog or induction resistivity (RDEEP), compressional slowness (DTP).In Uzen f ield, various types of neutron logging were recorded: 

• Neutron-gamma logging (NGK), which measures gamma radiation induced by neutrons on the rock. 

• Thermal neutron logging and epithermal neutron logging (NNKT), measuring the density of thermal and epithermal neutrons respectively. Initially, some data lacked the hydrogen index or neutron porosity curves, which were prepared for subsequent processing. Many wells only had raw neutron measurements, including count rates from far and near sondes or a neutron-gamma detector. Different formulas were applied to calculate neutron porosity, considering well diameter, drilling fluid properties, and reservoir conditions, based on the type of tools used. 

As the data spanned 60 years, information on well and mud data wasn’t consistently stored and accessible, necessitating the use of general formulas for recalculation, resulting in deviations in neutron porosity values in the reference zone. The choice of representative zones and similar characteristics was based on knowledge of the local reservoir geology and petrophysics: the average value of neutron porosity of clean sandstone in the Valanginian (12th horizon) is approximately 27% (first reference interval) and reading of Oxfordian clay (13th horizon) is around 42%  (second reference interval). Gamma ray logs were recorded using different tools and units of measurement, prompting the conversion of uR/hr to GAPI for further processing. Spontaneous potential logs were baseline shifted to consistent shale baselines with depth and were also converted to Alpha SP ranging from 0 to 1 to facilitate their use as an input for machine learning. 

Resistivity data appeared in different types: Russian laterolog (BK), Gradient resistivity (BKZ), array induction (VIKIZ). The traditional interpretation methodology of BKZ tools involved a complex set of chart-book lookup procedures to correct each curve for mud and wellbore properties, as well as for the effects of nearby low resistivity layers, thin beds, and large invasion zones. This manual procedure was challenging to apply, necessitated extensive interpolation, and typically provided only a single average value for each zone. Therefore, the authors opted to utilize laterolog resistivity for testing the machine learning algorithm in the Uzen field.