Data-driven modeling of time-domain induced polarization

Induced polarization (IP) measurements are affected by various types of noise, which should be removed prior to data interpretation. However, existing processing methods often rely on empirical assumptions about the standard shape IP decay curves. Our goal is introduce a data-driven approach for modeling and time-domain measurements. To reach this goal, we train variational autoencoder (VAE) 1,600,319 decays collected in Canada, United States, Kazakhstan. The proposed deep learning unsupervised avoids pitfalls parameterization with Cole-Cole Debye decomposition models, simple power-law or mechanistic models. Four applications VAEs key data: (1) synthetic generation, (2) Bayesian denoising, (3) evaluation signal-to-noise ratio, (4) outlier detection. Furthermore, interpret compilation’s latent representation reveal correlation between its first dimension average chargeability. Finally, determine that single real-valued scalar parameter contains sufficient information encode data. This new finding suggests using mathematical models governed more than one free ambiguous, whereas only chargeability justified. A pretrained implementation VAE model available as open-source Python code.