Physics-guided neural network for velocity calibration using downhole microseismic data

Hongliang Zhang, Jubran Akram, Kristopher A. Innanen

We present an unsupervised physics-guided neural network to calibrate the simplified 1D layered velocity model based on downhole recordings of perforation shots. This novel neural network incorporates five fully connected layers, a Scaling & Shifting layer as well as a forward modeling layer that generates theoretical travel times of P- and S-waves. Due to the inclusion of the forward modeling layer, our network eliminates the need for labeled data which is unavailable or limited in many cases. In addition, compared with conventional theory-based inversion, the neural network can solve the velocity optimization problem without explicit programming. To yield better constraint for both velocity-calibration and event-location problems, a hybrid objective function is used, which combines misfits of both arrival times and arrival-time difference between P- and S-waves. We apply the proposed neural network to a numerical example with six simulated perforation shots, yielding robust inversion results for layer velocities in the presence of noise. This neural network will be further examined with field data in the future research.