New VAE-Driven Inversion Method Enhances Fault Slip Prediction in CO₂ Storage Simulations

Global: New VAE-Driven Inversion Method Enhances Fault Slip Prediction in CO₂ Storage Simulations

A team of scientists has introduced a variational autoencoder (VAE)–based data‑space inversion (DSI) framework that directly infers pressure, stress, strain fields and fault‑slip tendency for carbon‑capture and storage projects. By leveraging thousands of prior geomodel simulations, the approach bypasses the need to generate posterior models while still delivering calibrated predictions from observed well data.

Challenges with Conventional History Matching

Traditional history‑matching techniques rely on iterative generation of posterior geomodels calibrated to field observations, a process that becomes computationally prohibitive in coupled flow‑geomechanics scenarios involving faults. The high dimensionality of permeability, porosity, and mechanical parameters often limits the practicality of these methods for large‑scale CO₂ sequestration studies.

Variational Autoencoder‑Based Data‑Space Inversion

The newly proposed workflow employs a VAE with stacked convolutional long short‑term memory layers to learn a compact latent representation of the simulated fields. Once trained on O(1000) prior geomodel outputs, the VAE enables rapid posterior inference by mapping observed pressure and strain data onto the latent space, eliminating the need for additional forward simulations.

Synthetic 3D Test Bed

Researchers evaluated the framework on a synthetic three‑dimensional model containing two faults. Heterogeneous permeability and porosity fields were generated via geostatistical software, while uncertain geomechanical and fault parameters were sampled from predefined prior distributions for each realization. Coupled flow‑geomechanics simulations were performed using the GEOS platform.

Training and Latent Variable Representation

The VAE was trained on the ensemble of prior simulations to encode pressure, strain, effective normal stress and shear stress fields into a set of latent variables. This parameterization captures the essential spatial patterns while reducing the dimensionality of the inversion problem, allowing the DSI step to focus on updating the latent variables in light of observed data.

Performance and Uncertainty Reduction

When applied to synthetic “true” models, the VAE‑DSI system produced accurate posterior predictions for pressure, strain, stress and fault‑slip tendency. Moreover, the method demonstrated a measurable reduction in uncertainty for key geomechanical and fault parameters, highlighting its potential to improve risk assessments in CO₂ storage operations.

Implications for Carbon Sequestration

By streamlining the inversion workflow and delivering reliable fault‑slip forecasts, the approach offers a scalable tool for evaluating the integrity of subsurface storage sites. Its ability to incorporate real‑time monitoring data could support more informed decision‑making and enhance the safety of large‑scale carbon capture initiatives.

This report is based on information from arXiv, licensed under Academic Preprint / Open Access. Based on the abstract of the research paper. Full text available via ArXiv.