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