From Simulation to Prediction: Data-Driven Modeling of Fluid Dynamics Using Artificial Intelligence

Author: Debabrata Pruseth
Publication Date: 2026/04/29
Document Type: Technical Note / Research Article
Language: English

Abstract

Computational Fluid Dynamics has traditionally relied on numerical methods to approximate the governing equations of fluid motion. These approaches remain foundational in engineering, atmospheric science, aerospace research, energy systems, materials science, and many other domains where the evolution of physical systems must be understood over time. However, traditional simulation approaches can become computationally demanding when physical systems involve high spatial resolution, long time horizons, turbulence, nonlinear interactions, or repeated scenario testing. Recent progress in scientific machine learning has introduced an alternative paradigm: instead of solving the governing equations explicitly for every new scenario, machine learning models can be trained to learn spatiotemporal patterns from physical data and generate predictions of future system states.
This paper presents a research-style study of a data-driven fluid prediction experiment using WALRUS, a cross-domain physics foundation model developed for continuum dynamics. The author implemented a workflow that uses WALRUS to predict the temporal evolution of Rayleigh–Bénard convection, a canonical heated-fluid system in which a fluid layer is heated from below and cooled from above. The implementation uses a structured physical field as input rather than a conventional image or video. The model receives a current state of the system and generates future states through autoregressive rollout, where each predicted state is recursively fed back into the model to produce subsequent predictions.
The study is grounded in the author’s blog article and GitHub notebook implementation. The blog frames the experiment as a shift from equation-first simulation toward datadriven prediction, while the GitHub notebook documents a practical implementation using Google Colab, PyTorch, WALRUS model checkpoints, and a Rayleigh–Bénard dataset. The paper discusses the methodology, model pipeline, qualitative observations, limitations, and future research directions. It also positions the experiment within the broader scientific machine learning landscape, including physics-informed neural networks, Fourier Neural Operators, machine-learning-accelerated CFD, and foundation models for physical systems.
The findings suggest that AI-based models can provide a useful complementary approach to traditional fluid simulation by learning approximate physical dynamics from data. However, the study also highlights important limitations, including autoregressive error accumulation, reduced interpretability, lack of explicit conservation-law enforcement, and the need for quantitative validation against numerical solvers.

Keywords
Artificial intelligence; fluid dynamics; computational fluid dynamics; WALRUS; Rayleigh–Bénard convection; scientific machine learning; surrogate modeling; autoregressive rollout; neural operators; Physics-Informed Neural Networks.

Download Research PDF
From Simulation to Prediction: Data-Driven Modeling of Fluid Dynamics Using Artificial Intelligence.

Suggested Citation
Pruseth, D. (2026). From Simulation to Prediction: Data-Driven Modeling of Fluid Dynamics Using Artificial Intelligence. Debabrata Pruseth AI Blog.

Companion Note
This page provides the abstract and full-text PDF for the research version of the article. A companion blog post explains the same work in a more narrative and implementation-focused style.

Read the companion blog:
https://debabratapruseth.com/from-simulation-to-prediction-learning-fluid-dynamics-with-ai/