Firstling digital twin: Data-driven time-space forecasting and simulation of vortex-induced vibrations
Name
ament-mentzelopoulos-phd-meche-thesis-2026.pdf
Size
30.74 MB
Format
Adobe PDF
Checksum (MD5)
0644d8fc93af953539986bcb33b14918
Author(s)
Mentzelopoulos, Andreas P.
Advisor(s)
Triantafyllou, Michael S.
Sapsis, Themistoklis
Date Issued
February 2026
Publisher
Massachusetts Institute of Technology
Abstract
We introduce the Firstling, a deep learning–based digital twin for flexible cylinders undergoing vortex-induced vibrations (VIV). The Firstling unifies three core capabilities: VIV motion reconstruction, forecasting, and simulation. Our first contribution is the development of data-driven augmentation strategies for improving physics-based VIV prediction models. We focus on enhancing the semi-empirical solver VIVA [106] through two complementary approaches. The first involves optimization-based refinement of the solver’s internal coefficient databases [52], achieving robust domain adaptation across benchmark VIV datasets [57]. The second applies deep learning–based correction layers on top of the solver, delivering improved performance and strong generalization across VIV regimes. Our second contribution is a systematic framework for riser motion reconstruction from experimental strain measurements. We demonstrate this approach on the MIAMI II experiments [87], providing the first published reconstruction of riser motions from this canonical dataset [53]. We formulate a constrained optimization problem, where a stochastic search algorithm addresses variable subset selection to identify optimal, physically consistent VIV motions from strain measurements. Third, we introduce the VIVformer [56], a transformer-based neural architecture for time-space reconstruction and forecasting of VIV. Trained extensively on both monochromatic and multi-frequency data, the VIVformer exhibits strong generalization even to unseen flow conditions. Its accuracy is validated using laboratory-scale data from the MIT Towing Tank and real offshore field measurements of a deepwater riser operating in the Gulf of America. Finally, we introduce the Virtual Towing Tank, a deep generative model for simulating VIV at arbitrary flow conditions. This development is motivated by realizing that a forecasting model trained on synthetic data produced by a variational autoencoder (VAE) can accurately predict unseen experimental VIV responses [55]. To identify the most physically consistent generative approach, a systematic comparison is conducted between generative adversarial networks (GANs), VAEs, and denoising diffusion probabilistic models (DDPMs) [54], demonstrating that DDPMs achieve the highest fidelity in replicating experimental statistics and dynamics. Building on this result, a conditional latent diffusion model is developed, capable of generating densely sampled and long time-horizon VIV data conditioned on flow velocity and tension.
MIT Department
Massachusetts Institute of Technology. Department of Mechanical Engineering
Massachusetts Institute of Technology. Center for Computational Science and Engineering
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link
