![[2601.13190] LAViG-FLOW: Latent Autoregressive Video Generation for Fluid Flow Simulations](https://arxiv.org/static/browse/0.3.4/images/arxiv-logo-fb.png){kind=logo}
[2601.13190] LAViG-FLOW: Latent Autoregressive Video Generation for Fluid Flow Simulations
Summary
LAViG-FLOW introduces a novel framework for generating fluid flow simulations, significantly improving efficiency and consistency in modeling subsurface multiphase fluid dynamics.
Why It Matters
This research addresses the high computational costs associated with traditional fluid flow simulations, which are crucial for applications like CO2 sequestration and geothermal energy. By leveraging a latent autoregressive video generation approach, LAViG-FLOW offers a faster and more efficient alternative, potentially transforming practices in environmental engineering and resource management.
Key Takeaways
- LAViG-FLOW utilizes a latent autoregressive model for fluid flow simulation.
- The framework achieves two orders of magnitude faster performance than traditional methods.
- It maintains consistency in saturation and pressure fields over time.
- The model is trained on a CO2 sequestration dataset, demonstrating practical applicability.
- This approach could lower costs and improve safety in subsurface fluid management.
Computer Science > Machine Learning arXiv:2601.13190 (cs) [Submitted on 19 Jan 2026 (v1), last revised 16 Feb 2026 (this version, v2)] Title:LAViG-FLOW: Latent Autoregressive Video Generation for Fluid Flow Simulations Authors:Vittoria De Pellegrini, Tariq Alkhalifah View a PDF of the paper titled LAViG-FLOW: Latent Autoregressive Video Generation for Fluid Flow Simulations, by Vittoria De Pellegrini and Tariq Alkhalifah View PDF HTML (experimental) 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 to 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, LAV...
