Research

Graph-based surrogate models offer a promising route to accelerate computational fluid dynamics (CFD) simulations on unstructured meshes. However, their development is limited by the scarcity of benchmark datasets spanning multiple flow regimes and standardized protocols for long-horizon autoregressive prediction. We introduce ONECYL (ONE CYLinder), a new benchmark for unsteady flow past a circular cylinder across laminar, transitional, and high-Reynolds-number regimes. The benchmark comprises 450 high-fidelity Variational Multiscale finite-element simulations (270,000 flow snapshots) with randomized cylinder geometries, providing time-resolved velocity and pressure fields together with mesh connectivity, geometric descriptors, Reynolds numbers, and integrated aerodynamic quantities. Beyond the dataset, ONECYL establishes a unified evaluation framework combining full-field rollout errors, virtual probes, and drag and lift predictions to assess numerical accuracy and physical fidelity. To accompany the benchmark, we develop a Graph Transformer as a reference baseline predicting velocity and pressure fields autoregressively on unstructured meshes. Using ONECYL, we investigate geometric representations and physics-based regularization across the three Reynolds-number regimes. The results show that explicitly encoding the cylinder geometry through a level-set representation consistently improves long-horizon prediction accuracy and generalization to unseen geometries, while divergence-based regularization becomes increasingly beneficial as flow complexity increases. The ONECYL benchmark and its Graph Transformer baseline provide a reproducible framework for evaluating graph-based surrogate models and establish a foundation for future research on long-horizon prediction of unsteady bluff-body flows.

This thesis presents a first major step towards the development of digital twins for industrial-scale solar power plants, with a focus on agrivoltaic systems. As recent innovative solar energy systems face increasing pressures from environmental hazards such as windstorms, there is a strong need for developing rapidly-assessed models (such as digital twins) of real-world power plants to support industrial decision-making in design, monitoring, and operation. Such models would contribute to the development of renewable energy technologies tailored to site conditions and operational requirements, which applied to agrivoltaics would ultimately accelerate Europe's green energy transition. Our work proposes a comprehensive and adaptable framework, that combines high-fidelity computational fluid dynamics (CFD) simulations, adaptive meshing, deep reinforcement learning (DRL) optimization techniques, and graph neural network (GNN) surrogate modeling to accurately simulate, optimize, and predict the performance of solar panel arrays under various wind conditions. This work, along with discussions on the remaining steps to achieve a full digital twin with in-situ feedback loops, shows the feasibility and potential of using advanced numerical methods and machine learning approaches to create digital twins of solar power plants.

Modeling the complex aerodynamics of solar power plants with precision is crucial for ensuring their safety in wind flow. Factors like scale, positioning, and wind exposure significantly affect the complexity of simulations, making it more challenging to accurately capture the behavior of turbulent wind interactions in large-scale panel arrays. In power plants using solar trackers, the variable tilt of the panels and fluctuating wind conditions further complicate the problem, especially when time-dependent and directional wind changes are considered. While Computational Fluid Dynamics (CFD) simulations offer high-fidelity modeling, their computational cost increases significantly with the mesh resolution required for large-scale configurations, limiting their feasibility in industrial applications. To overcome these computational challenges, we apply advanced machine learning methods, specifically Graph Neural Networks…

Although a key driver for green energy development, solar photovoltaic power plants face the major risk of severe wind damages, as there is currently no best practice on how to best stow the panels under strong wind conditions. In this research, an out-of-the-box numerical framework is introduced to inform the discussion around panel design and recommended stow positions, one that leverages the increasing use of solar tracker actuators, that allows panels to set an optimal angle relative to the sun to maximize power output, and incidentally offer a great potential for optimal safeguarding through individual panel piloting. The task of concurrently optimizing multiple panel tilts in a turbulent atmospheric boundary layer wind flow is modeled as a Markov decision process and solved with a single-step deep reinforcement learning algorithm, intended for situations where the optimal policy to be learnt by a neural network does not depend on state. The numerical reward fed to the neural network is computed from high-fidelity numerical simulations combining variational multiscale modeling of the Navier–Stokes equations and anisotropic boundary layer mesh adaptation, to accurately represent critical flow features at affordable computational costs, regardless of the panel tilts chosen by the learning agent. A range of experiments is performed across various learning objectives accounting for different possible causes of breakage (such as tear, vibrations, and fatigue), for which the proposed approach successfully minimizes the aerodynamic efforts on two-dimensional and three-dimensional arrangements of six ground-mounted panels under an incident wind speed of 50 km/h, while outperforming baseline safeguarding practices considered in the literature by several dozen per cent. This gives hope that, by interacting with its computational fluid dynamics environment in a trial-and-error manner, a deep reinforcement learning agent can learn unexpected solutions to this complex decision-making problem and come up with innovative, feasible solutions capable of managing utility-scale solar assets during high-wind events while efficiently complementing engineering intuition and practical experience.

Prompt ElastoGravity Signals are light-speed gravity-induced signals recorded before the arrival of seismic waves. They have raised interest for early warning applications but their weak amplitudes close to the background seismic noise have questioned their actual potential for operational use. A deep-learning model has recently demonstrated its ability to mitigate this noise limitation and to provide in near real-time the earthquake magnitude (Mw). However, this approach was efficient only for large earthquakes (Mw ≥ 8.3) of known focal mechanism. Here we show unprecedented performance in full earthquake characterization using the dense broadband seismic network deployed in Alaska and Western Canada. Our deep-learning model provides accurate magnitude and focal mechanism estimates of Mw ≥ 7.8 earthquakes, 2 minutes after origin time (hence the tsunamigenic potential). Our results represent a major step towards the routine use of prompt elastogravity signals in operational warning systems, and demonstrate its potential for tsunami warning in densely-instrumented areas.