Compute-driven use cases towards Exascale
One of the primary goals in CoE RAISE is the development and expansion of AI methods along representative use cases from research and industry, which have a strong focus on data-driven technologies, i.e., analyzing data-rich descriptions of physical phenomena. Example use cases vary widely and range from fundamental physics and remote sensing to 3D printing and acoustics.
Use Case 1: AI for turbulent boundary layers
Active Drag Reduction (ADR) of Turbulent Boundary Layer (TBL) flows is the focus of this first task of WP3, where the multi-physics code multiphysics Aerodynamisches Institut Aachen (m-AIA), written in C++, is used to predict the friction drag of TBL flow subjected to spanwise traveling transversal surface waves, which are known to reduce the friction drag. These numerical computations and the obtained data build the foundation of the subsequent modeling approaches, which in turn are expected to predict even better
actuation principles.
Recent results have demonstrated that, for large enough wavelengths, a good scaling of the skin friction reduction is obtained thanks to the actuated surface waves. Figure 1 shows an instantaneous result of this actuation, with the left (non-actuated) surface having much higher activity than the right (actuated) one.
Figure 1: Contours of positive and negative near-wall velocity streaks in the near-wall region for (left) a
non-actuated flat plate boundary layer flow and (right) an actuated boundary layer flow with high drag
reduction; red: high speed streaks (𝑢′+ = +3); blue: low speed streaks (𝑢′+ = −3).
An AutoEncoder (AE) neural network was then set up and trained to reproduce 3D near-wall flow fields, and compared with a novel approach named a Physics-Constrained AutoEncoder (PC-AE). Preliminary results show a better quantitiative agreement for the simple AE for now, and further investigation is needed to explore how to improve the inclusion of needed physical constraints in this dimensionality reduction technique.
Use Case 2: AI for wind farm layout optimization
The second use case is investigated jointly by the Barcelona Supercomputing Center (BSC) and its AI collaborator, Riga Technical University (RTU). It focuses on the the wind energy industry, where simulations of wind farms are required at the design stage and also for short-term (days) and long-term (months to years) power predictions. At the design stage, the objective is to assess the potential of the farm and to optimize the placement of the wind turbines on the terrain. Short- and long-term forecasts are necessary for the operator to assess the wind resources available in the coming days or months. In all cases, an accurate description of the airflow through topography, generally involving complex and heterogeneous terrains, and the wind turbine influence on the airflow are necessary.
Solving for the flow over a full turbine farm requires reduced order models for the actual turbines, as there are too many scales between the dynamics on the turbine blades and the full farm for a fully resolved simulation. In RAISE, the objective is to model the turbines using a deep learning procedure, replacing the turbine by an equivalent “sink” for the Navier-Stokes equations. This algorithm is trained on detailed simulations run with BSC’s Alya solver, a strategy described in Fig. 2.
Figure 2: Training process to create the wind turbine surrogate.
Recently, simple initial neural networks were trained on this task recently, using the PyTorch framework. Specifically, an approach known as RBF-RNN was implemented, with the specific constraint that the architecture must be implemented in a highly scalable manner inside Alya once the training is done. The workflow is now functional for a simple model, proving the concept and paving the way for more elaborate tests and approaches in the near future.
Use Case 3: AI for data-driven models in reacting flows
This use case focuses on another growing challenge related to energy transition: the simulation of combustion of new, carbon-free fuels. Indeed, turbulent combustion is notoriously difficult to model, and the use of fuels such as Hydrogen, expected to become an important energy vector in the coming decades, will dramatically impact existing models that were designed for fossil fuels. This task suggests focusing on hydrogen and trying to desing new “subgrid-scale” (SGS) models for flame-turbulence interactions.
Highly-resolved H2 combustion simulations were run, and filtering procedures used to produce a training database (Fig. 3). Neural networks are then trained for the task of predicting subgrid-scale flame-turbulence interactions from this data. A major diffculty, analogous to use case 2, arises when trying to use these networks inside AVBP-DL, a running high-fidelity solver. Preliminary results have shown however, that several strategies, based on explioting hybrid CPU/GPU hardware, led to good scaling of the problem,
paving the way for simulations with deep learning-based SGS models in 2022.
Figure 3: H2 mass fraction of a 3D DNS with temporal decaying HIT. The process of creating a database for
SGS models by filtering and downscaling a DNS is shown from top to bottom
Use Case 4: AI for next-generation aircraft engine design
SAFRAN, a world leader in designing, building and maintaining aircraft engines, is actively exploring new designs for engines based on carbon-free tools, among which hydrogen is an excellent candidate. In this use case, the aim is to explore the challenges of designing an H2-based helicopter engine. H2 has very different characteristics compared to the kerosene that aircrafts typically burn, and the combustor designed must be entirely reimagined.
In this use case, the AVBP solver is adapted for H2 combustion. To this end, a reduced chemical kinetic scheme was designed to describe H2 combustion at a reduced cost compared to fully detailed schemes, while retaining most of the flame’s structure (Fig. 4). High-fidelity simulations of turbulent reactive flows were then produced, feeding use case 3 with data for SGS model approaches. These encouraging results have prepared the ground for the next steps of the project, where more realistic simulations will progressively lead to the simulation of a full H2 combustor for a helicopter engine.
Use Case 5: AI for wetting hydrodynamics
The last use case investigates wetting hydrodynamics, i.e. dynamics of liquids in contact with surfaces. It is lead by the Cyprus Institute (CYI) and the University of Iceland (UOI). Modeling of this phenomenon requires modelling of multiple physical interactions, at disparate scales. This challenge, shared with all use cases of WP3, can be intractable in practical applications, such as novel technologies in microfabrication, the biomedical, smart materials, pharmaceutical and printing industries, as well as energy conversion and water harvesting in environmental applications, to name a few. To circumvent these limitations, researchers
Figure 4: 1D Flames with detailed (black) and simplified (“skeletal”, red) kinetic schemes.
develop surrogate models for parts of the problem, avoiding the direct resolution of part of the physics. This is a difficult, mostly intuition-driven task however, and in RAISE the aim is to explore the use of data-driven techniques instead.
The process of developing data-driven surrogate models for wetting hydrodynamics involves (i) the generation of datasets that are sufficiently diverse to represent a breadth of wetting scenarios on chemically heterogeneous surfaces, e.g., striped features, randomly distributed features, chemical gradients etc., and (ii) the training of AI models that are expressive enough to learn the complex physical mechanisms that drive droplet transport.
In the first year, improvements to the CFD solver used to produce the training databases has been performed, as well as careful validation of the results. This solver was used to produce a high-fidelity database, com- plemented with a larger, lower-quality database produced using a low-order model. A data-driven strategy based on Fournier Neural Operators was then implemented, and trained to predict the evolution of the wetted surface over time. Figure 5 shows results from all of these approaches.
Figure 5: (𝑎) Snapshot of the wet surface of a spreading droplet on a striped surface, during the final time- step of the simulation using the low-order model. (𝑏) Comparison between the low-order model and the one-dimensional FNO prediction. (𝑐) Comparison between the low-order model and the two-dimensional FNO prediction
Upcoming work will focus on optimizing and diversifying the neural network approach, and increase the database size by running larger simulations with the validated CFD code.
What’s next for WP3?
WP3 brings together varied topics from the computational fluid dynamics (CFD) domain: external aero- dynamics for drag reduction and electricity production, reactive flows for H2 combustion, and wetting hy- drodynamics. In all of these, challenging applications, spanning multiple physics and large scales of length and time are identified. These cannot be directly fully resolved, even using very high performance exascale computations. The new challenge is therefore to seek strategies to supplement existing physical solvers with data-driven approaches, leveraging the strengths of each approach, and complementing the weaknesses of the other.