Virtually any property that is logged or measured in the wellbore, e.g., velocities, derivative attributes, Gamma Ray, Resistivity, Density, Stresses, TOC, Brittleness, Pore Pressure, etc. Also drilling measurements such as UCS, wellbore stability, WOB, etc.

Seismic (stack and angle stack reflectivities and derived attributes) and interpreted surfaces. Well logs and well tops for training.

If you provide # of wells with full log suites, areal extent of 3D seismic and number of properties you wish to have simulated, we can provide indicative pricing. Submit a Technical Presentation Request if interested in learning more

There’s generally not enough data to justify the use of a Deep Learning neural network so we generally use a simple ANN. Where the “secret sauce” comes into play is how one sets up the NN to reference the data properly. We don’t just look at one data point at the trace at a particular depth – we look at a user-specified number of samples within a sliding/overlapping reference window.

We do not explicitly extract the seismic wavelet. We treat the seismic trace as a time (or depth) series and sample it at a resolution that approaches log scale.

Much of the QEarth process is automated. In addition, time consuming tasks such as rock physics modeling, wavelet extraction and low frequency model building are not performed in this data-driven approach. Hence the dramatic reduction in time-to-solution.

A) Sub-salt GOM and pre-salt Brazil, offshore Malaysia, offshore onshore Alaska, onshore Middle East, offshore Oman, onshore Colombia. B) Several in the Permian Basin and Eagle Ford in Texas. SCOOP/STACK play in Oklahoma. Results have mostly been good to excellent. Only one was disappointing due to poor well availability and imaging issues (poor seismic-to-well tie)

The low frequency trend/depth dependency is explicitly present in the input log data and is "learned" by the neural network during the training phase.

We use a process called Bagging (Bootstrap aggregation) where the algorithm randomly selects 20% - 30% of the input data and uses it to verify simulations. We also randomly withhold wells from the training set in a round-robin fashion. This is called Grid Search with K-Fold cross-validation.

We are quite flexible in terms of the input seismic attributes. Effectively we can use the same attributes as a standard seismic inversion, but we are not limited to amplitude data. We can even input AVO attributes and conventional inversion attributes. One key point to make is that we can explicitly invert for properties rather than deriving them mathematically from Ip, Is, and Density. This is especially important when one considers we can use QEarth to invert for drilling dynamics data, such as weight on bit, pore pressure, fracture gradient, wellbore stability and UCS, etc.

No. It stands to reason that an inversion model would benefit from a larger number of wells, but over-training may happen. Depending on the size of the volumes and attributes of the data varying populations of wells are used.

Neural networks are not great extrapolators but Quantico will conduct extensive analysis of well control prior to kick-starting a project. Before a project starts, we will communicate concerns about poorly sampled areas of a volume and determine what KPIs are feasible given the data available.