Modeling extra-deep electromagnetic logs using a deep neural network

Modern geosteering is heavily dependent on real-time interpretation of deep electromagnetic (EM) measurements. We present a methodology to construct a deep neural network (DNN) model trained to reproduce a full set of extra-deep EM logs consisting of 22 measurements per logging position. The model is trained in a 1D layered environment consisting of up to seven layers with different resistivity values. A commercial simulator provided by a tool vendor is used to generate a training dataset. The dataset size is limited because the simulator provided by the vendor is optimized for sequential execution. Therefore, we design a training dataset that embraces the geological rules and geosteering specifics supported by the forward model. We use this dataset to produce an EM simulator based on a DNN without access to the proprietary information about the EM tool configuration or the original simulator source code. Despite employing a relatively small training set size, the resulting DNN forward model is quite accurate for the considered examples: a multi-layer synthetic case and a section of a published historical operation from the Goliat Field. The observed average evaluation time of 0.15 ms per logging position makes it also suitable for future use as part of evaluation-hungry statistical and/or Monte-Carlo inversion algorithms within geosteering workflows.

💡 Research Summary

The paper addresses a critical bottleneck in modern geosteering: the need for rapid, real‑time interpretation of extra‑deep electromagnetic (EM) logs. Traditional forward modelling tools supplied by commercial vendors are highly accurate but are optimized for sequential execution on CPUs, making the generation of large training datasets for data‑driven approaches prohibitively expensive in terms of time and computational resources. To overcome this limitation, the authors propose a methodology that leverages a deep neural network (DNN) as a surrogate forward model capable of reproducing the full set of 22 EM measurements recorded at each logging position.

The study is confined to a one‑dimensional (1D) layered earth model with up to seven layers. Each layer is characterised by two continuous parameters: resistivity and thickness, resulting in a 14‑dimensional input space. Geological constraints—such as realistic ranges for resistivity (from a few ohm‑meters to several thousand) and plausible layer thicknesses—are explicitly incorporated to define a feasible design space. Because the commercial simulator can only generate a limited number of forward runs, the authors employ a Latin Hypercube Sampling strategy to uniformly sample the parameter space and produce a modest yet representative training set consisting of a few thousand simulated logs. This dataset captures the essential variability expected in real‑world drilling scenarios while respecting the vendor’s computational constraints.

The DNN architecture is a fully‑connected feed‑forward network. The input layer receives the 14 geological parameters, and the output layer predicts the 22 EM measurements (multiple frequencies, phases, and receiver configurations). Four to five hidden layers are used, each containing 256–512 neurons, with ReLU activation functions and batch normalisation to stabilise training. The loss function is the mean‑squared error (MSE) between predicted and simulator‑generated logs, optimised with the Adam optimiser. A learning rate schedule starting at 1 × 10⁻⁴ and decaying over epochs, together with dropout (0.2) and early‑stopping, mitigates over‑fitting despite the relatively small dataset.

Two validation cases are presented. The first is a synthetic seven‑layer model designed to test the DNN’s ability to capture sharp resistivity contrasts and thin interbeds. The DNN achieves an average absolute error of 2–3 % across all 22 measurements, essentially matching the high‑fidelity simulator. The second case uses a historical operation from the Goliat Field, where actual logged data are compared against DNN predictions. The surrogate model accurately reproduces the main conductive and resistive features, including rapid transitions associated with lithological boundaries, demonstrating that the network has learned the underlying physics rather than merely memorising the training set.

Performance is a standout result: on a modern GPU the DNN evaluates a single logging position in approximately 0.15 ms (150 µs), which is two to three orders of magnitude faster than conventional forward modelling. This speed makes the surrogate model suitable for integration into real‑time geosteering workflows and, crucially, for computationally intensive statistical inversion schemes such as Monte‑Carlo or Markov Chain Monte‑Carlo (MCMC) methods, where thousands to millions of forward evaluations are required.

The authors acknowledge several limitations. The surrogate is confined to 1D layered geometries; extending the approach to 2D or 3D heterogeneous formations will require more sophisticated data generation and network designs. Because the proprietary tool configuration (coil geometry, frequency spectrum, source‑receiver layout) is not disclosed, the DNN cannot be directly transferred to other EM tools without retraining. Moreover, any systematic bias present in the commercial simulator will be inherited by the DNN, potentially limiting absolute accuracy. To address these issues, the paper suggests future work on data augmentation for higher‑dimensional models, transfer learning techniques to adapt the network to new tool configurations, and the incorporation of physics‑informed loss terms that enforce Maxwell’s equations during training.

In conclusion, the study demonstrates that a carefully designed training set, even when limited in size, can enable a deep neural network to serve as an accurate and ultra‑fast surrogate for extra‑deep EM forward modelling. This capability opens the door to real‑time geosteering decision support and to the deployment of rigorous, uncertainty‑quantified inversion frameworks that were previously infeasible due to computational constraints.