VIVALDy: AI-Driven Low-Order Modeling of Vortex-Induced Vibrations via β-Variational Autoencoders, Transformers, and Adversarial Training

Abstract

This paper presents VIVALDy, a machine-learning framework for low-order modeling of vortex-induced vibrations (VIV). The approach combines a β-Variational Autoencoder-Generative Adversarial Network (β-VAE-GAN) with masked convolutions to extract compact, interpretable latent representations of the turbulent flow field around a moving cylinder, and a bidirectional transformer to model the temporal evolution of these features from minimal sensor measurements. Validated against experimental data across multiple fluid-structure interaction regimes, the framework accurately reconstructs turbulent flow states using only cylinder displacement as input.

Paper

Recommended citation: Tonioni, N., Agostini, L., Kerhervé, F., Cordier, L., & Vinuesa, R. (2025). VIVALDy: AI-Driven Low-Order Modeling of Vortex-Induced Vibrations via β-Variational Autoencoders, Transformers, and Adversarial Training. AiFluids 2025, Chania, Greece.
Download Paper