LAViG-FLOW: Latent Autoregressive Video Generation for Fluid Flow Simulations

TL;DR

The paper addresses the computational burden of forecasting coupled subsurface multiphase flow fields by introducing LAViG-FLOW, a latent autoregressive video diffusion framework. It uses dual autoencoders (2D VQ-VAE for saturation and 2D VAE for pressure) to compress state variables into latents that feed a Video Diffusion Transformer (ViDT), which learns their joint spatiotemporal distribution. Through autoregressive fine-tuning, the model extrapolates beyond the training horizon while maintaining physical consistency, delivering orders-of-magnitude speedups over traditional reservoir solvers on a CO$_2$ sequestration dataset. The approach is flexible, scalable to additional variables, and validated with both qualitative and quantitative analyses, with code and data release planned.

Abstract

Modeling and forecasting subsurface multiphase fluid flow fields underpin applications ranging from geological CO2 sequestration (GCS) operations to geothermal production. This is essential for ensuring both operational performance and long-term safety. While high fidelity multiphase simulators are widely used for this purpose, they become prohibitively expensive once many forward runs are required for inversion purposes and quantify uncertainty. To tackle this challenge we propose LAViG-FLOW, a latent autoregressive video generation diffusion framework that explicitly learns the coupled evolution of saturation and pressure fields. Each state variable is compressed by a dedicated 2D autoencoder, and a Video Diffusion Transformer (VDiT) models their coupled distribution across time. We first train the model on a given time horizon to learn their coupled relationship and then fine-tune it autoregressively so it can extrapolate beyond the observed time window. Evaluated on an open-source CO2 sequestration dataset, LAViG-FLOW generates saturation and pressure fields that stay consistent across time while running orders of magnitude faster than traditional numerical solvers.

Figures (6)

• Figure 1: Overview of the LAViG-FLOW pipeline: A reservoir simulator provides $N$ video frames of CO2 gas saturation and corresponding pressure build-up fields, which are first compressed via 2D VQ-VAE and 2D VAE encoders, concatenated, patch-embedded, and passed through spatio-temporal Transformer blocks before being unpatchified; the VQ-VAE decoder recovers CO2 gas saturation and the VAE decoder recovers pressure build-up, so the pipeline generates $F{>}N$ video frames at inference.
• Figure 2: Training workflow for Stage I: the pressure build-up images are compressed into a continuous latent representation, while the CO$_2$ gas saturation images are compressed into a vector-quantized latent representation. Here, $\mathbf{x}$ denotes the input frames and $\hat{\mathbf{x}}$ denotes the reconstructed frames after decoding.
• Figure 3: Stage II (pre-training): Gaussian noise is used to corrupt the joint latent video and VDiT-Base removes it; at inference, VDiT-Base denoises Gaussian noise into an $N$-frame video (same length as input). Stage III (autoregressive fine-tuning): The same workflow is fine-tuned into VDiT-AR (autoregressive) to predict future frames from $N$ context frames, yielding $F{>}N$ frames; inference is unconditional (start from noise) or conditional (fix encoded context frames). Symbols: $\mathbf{x}$ is the input clip of length $N$ (context frames only, Stage II) or $N$ + target frames (Stage III); $\mathbf{z}$ concatenated latents; $\mathbf{z}_{t=T}$ fully noised latents at step $(t{=}T)$; $\tilde{\mathbf{z}}$ denoised latents; $\tilde{\mathbf{x}}$ decoded clip of length $N$ (Stage II) or $F{>}N$ (Stage III).
• Figure 4: 21 unconditionally generated videos from the VDiT model (Stage II). The top 21 rows show CO2 gas saturation field across 17 frames; the next 21 rows show the corresponding pressure build-up field, respectively.
• Figure 5: 19 unconditionally generated videos using the autoregressive strategy (Stage III). The top 19 rows show CO2 gas saturation across 23 frames; the next 19 rows show the corresponding pressure build-up, respectively. Labels for frames 18--23 (beyond the 17-frame training horizon) are shown in red.