Uncertainty in aeroelastic predictions forces aircraft manufacturers to rely on conservative design margins, increasing weight and certification costs. Using exascale fluid–structure simulations, CEEC demonstrated how high-fidelity modelling can improve load prediction, reduce development risk, and support more efficient aircraft designs.
| Industry | Scientific Domain | Code |
| Aerospace | Fluid–Structure Interaction (FSI) | Alya/SOD2D |
Description of the Challenge
In addition to studying the phenomenon of shock buffet in our first Lighthouse Case, certifying aircraft requires understanding its impact on wing structural dynamics, or how air forces move and stress the wing. Our second lighthouse case (LHC2) addresses this knowledge gap by performing a high fidelity aeroelastic simulation of the SFB 401 wing in a transonic regime (Ma = 0.8, just below the speed of sound) using existing large eddy simulation (LES) models for compressive flows. This case study is also known by the research community as the HIRENASD wing model.
The accuracy of aeroelastic predictions in this regime is crucial for certification-relevant quantities such as flutter margins, buffet onset, and dynamic load envelopes. However, standard industrial approaches based on RANS fluid solvers and linear structural models are not capable of resolving the unsteady, and by definition nonlinear, shock-driven flow physics that dominate near Ma = 0.8. As a result, key nonlinear fluid–structure interaction effects remain under-modelled.
These modelling limitations have direct practical consequences. In order to maintain certification safety, industry workflows often rely on conservative design margins, additional, time-consuming experimental wind tunnel campaigns, and structural over-sizing. Each of these measures adds significant cost, weight, and operational penalties to flying over the lifetime of the aircraft.
To overcome these modelling limitations, LHC2 implements a high-fidelity, fully coupled aeroelastic framework that can help industry reduce both development time and cost.
Why this Matters for Industry
The limitations of linearised aeroelastic methods, especially for transonic flight, are well understood by the engineering community. What has been missing is a computationally feasible path to something more accurate and realistic. LHC2 addresses this well-known knowledge gap directly:
- Improved prediction of flutter boundaries that are grounded in resolved shock–boundary layer physics rather than empirically padded safety margins, which enables structural mass reduction where current conservatism drives over-design.
- More accurate nonlinear load predictions reduce the number of wind tunnel tests required for aeroelastic certification, one of the most time-consuming and expensive stages of aircraft certification.
- Aircraft wing design can exploit combining different materials to tailor bending-torsion coupling and achieve desired stiffness.
These benefits depend strongly on the accuracy of the aerodynamic loads; therefore, high-fidelity coupled simulations become essential for a reliable design process.
Moreover, high-aspect-ratio and laminar-flow wings, which will be central to the next generation of short- and medium-range aircraft because of their potential for higher energy efficiency, operate in flight regimes where understanding nonlinear aeroelastic effects is particularly important. The methodology demonstrated in LHC 2 can also be applied to these configurations, thereby shortening development and certification timelines for aircraft using these newer classes of wings.
▶ Scientific Background
- understanding of non-linearities that appear in transonic regimes for real wing reference configurations,
- accurate determination of aeroelastic deformation in flight conditions,
- computational advances for coupled problems towards Exascale Computing.
Most of these gaps exist due to a lack of development of computational methods capable of solving these problems, as well as the high computational and monetary cost of the coupled simulations, which tend to be impractical in a production setting.
The HIRENASD wing model configuration provides experimental wind tunnel data against which the coupled simulation can be directly validated, giving the results quantitative credibility beyond what a purely computational study can claim. The same fluid–structure interaction (FSI) methodology was verified during development using the modified Turek–Hron benchmark, a controlled reference case for FSI solvers.
The information that this approach captures that linearised solvers do not includes: shock–boundary layer interaction, unsteady shock oscillations driving buffet, and the nonlinear geometric coupling between structural deformation and flow topology. These are not second-order effects in transonic aeroelasticity — they are the mechanisms that directly determine whether a certification prediction is conservative or unconservative.
▶ Technical Details: Exascale Computing Approach
Achieving the required fidelity for the coupled aeroelastic problem is not simply a matter of increasing computational resources within an existing workflow; it requires a redesign of how the multiphysics system is partitioned, coupled, and scaled in order to take advantage of modern supercomputers. Throwing more processors and ram at a problem only works when the software can efficiently divide its problem across those extra processors and ram.
Simulations operate at large scale, involving on the order of hundreds of millions of fluid nodes and millions of structural nodes. The workflow runs on distributed HPC architectures using both CPU and GPU resources, with several thousand cores depending on the configuration.
From a scalability perspective, both strong and weak scaling behaviour for LHC2 have been evaluated. Holding the complexity of the problem constant, time to solution is reduced in proportion to hardware allocated (strong scaling) or holding time constant, problem size can be increased as long as hardware allocation is also increased proportionally (weak scaling). The coupled toolchain maintains high parallel efficiency at large problem sizes, while the fluid solver alone demonstrates scalability to several thousand GPU nodes, indicating robust performance on extreme-scale computing architectures.
As a result of these software developments, simulations that were previously computationally prohibitive are now feasible within leadership-class HPC environments, enabling realistic high-fidelity analysis of complex transonic flow configurations and providing a pathway towards exascale-class applications.
Results and New Insights
The coupled simulations capture aeroelastic phenomena in transonic flow regimes that are not fully represented in uncoupled or linearised RANS-based approaches.
Based on the new coupled LHC2 simulation, we can now show that wing deformation under transonic aerodynamic loading can alter shock position along both chord and span. This interaction forms a strongly coupled nonlinear system in which structural deformation modifies the flow field, which in turn feeds back into the aerodynamic loads. Accurately resolving this feedback requires a two-way coupled fluid–structure interaction framework, as one-way or loosely coupled approaches may not correctly predict the resulting load redistribution, which is relevant for structural sizing and flutter margin assessment.
Furthermore, LHC2 results show that nonlinear phenomena such as shock-induced unsteadiness and amplitude-dependent damping, as well as modal coupling, emerge from the coupled fluid–structure system itself and are not captured by frequency-domain linearised methods, which are restricted by their underlying assumptions of small perturbations about a steady operating point.
The visualisation shows the local Mach distribution across the wing surface at transonic conditions: high Mach concentrated near the leading edge, transitioning sharply across the shock, with lower pressure extending toward the trailing edge and wing tip. The shock position and its spanwise variation are directly visible — features whose accurate prediction is at the core of what distinguishes high-fidelity aeroelastic simulation from standard industrial approaches.
On the computational side, performance optimisation of the solver and toolchain has reduced both time-to-solution and energy-to-solution across the tested configurations. Strong and weak scaling studies indicate that SOD2D maintains parallel efficiencies above approximately 75% up to 4,096 GPUs for the considered benchmark cases, demonstrating good scalability on current large-scale HPC systems and providing a basis for further extension towards emerging exascale architectures, subject to problem size and hardware characteristics.
The performance of Alya has been improved through optimizations of the assembly kernels in the mesh motion and structural mechanics modules. For the latter, a speedup of up to 7.6× has been achieved relative to the baseline implementation, significantly reducing both the time-to-solution and energy-to-solution for Fluid-Structure Interaction (FSI) simulations. In addition, a parallel efficiency of 95% has been demonstrated on up to 12,544 cores for FSI problems.
Industrial Takeaways
The value of LHC2 for industry is not primarily academic. It demonstrates a simulation capability that addresses specific, well-known deficiencies in current aeroelastic practice:
- Resolved nonlinear transonic aerodynamics provides load predictions that are physically better founded than RANS-based approaches, reducing the uncertainty that currently drives conservative structural sizing and over-design.
- The validated FSI methodology on the HIRENASD configuration establishes a credible basis for integration into industrial aeroelastic workflows, either as a high-fidelity reference tool or as a source of correction data for lower-fidelity models.
- Improved predictive confidence at transonic conditions directly reduces the wind tunnel test matrix required for flutter and buffet clearance, with measurable aircraft certification cost and schedule implications.
Even where full LES–FSI simulation remains too costly for routine production use, results from LHC2 can calibrate reduced-order models or inform machine learning surrogates that improve the fidelity of standard RANS-based pipelines — a pathway to industrial impact that leverages existing workflows instead of requiring their replacement.
▶ Codes and Software Stack
Both codes are developed at the Barcelona Supercomputing Center. SOD2D is open source; Alya is available under an academic licence. Their partitioned coupling architecture and reliance on established open standards (MPI, OpenACC, FEM/SEM) make the tool chain adaptable to industrial CFD/FEM ecosystems without requiring wholesale replacement of existing infrastructure.
Alya handles the structural mechanics and fluid–structure coupling. Its ALEFOR module manages mesh deformation under the ALE strategy, continuously updating the computational grid as the wing displaces. The structural assembly (SOLIDZ) and fluid assembly (NASTIN) modules run concurrently with the fluid solver on separate MPI partitions. During this phase, the mesh-movement kernels were explicitly vectorised and the mesh motion problem reformulated as three decoupled scalar Laplacian equations, delivering a 1.8× module-level speedup while preserving mesh quality and numerical accuracy. In terms of functionalities, two main lines have been followed. On the one hand, the generalization of spectral element methods (SEM), including: the integration of new elements; the generation of SEM meshes from linear ones (TET or HEX); the implementation of static condensation techniques to accelerate the solver convergence. On the other hand, multi-resolution coupling has been implemented to enable the use of non-conforming meshes using different orders of approximation and methods (finite element methods (FEM) and SEM). The methodology can be used for implicit and explicit couplings, and seamlessly to partitioned approaches (e.g. FSI).
SOD2D is a high-order spectral element CFD solver, GPU-accelerated via OpenACC and parallelised with MPI. It provides the fluid solution — resolving turbulent transonic flow on the deforming wing geometry — and supports compressible and incompressible regimes with multiple time-integration schemes (LSRK, IMEX-RK, BDF-EXT3). Kernel-level optimisations carried out in this phase, including loop reorganisation, collapse directives, improved memory coalescing, and reduction of kernel-launch overhead, substantially improved device utilisation and strong-scaling behaviour. Exascale readiness was demonstrated at 4,096 GPU nodes on MareNostrum 5.