The aeronautics industry is a sophisticated user of computational science and engineering (CSE), and a source of some of the field’s most difficult challenges. Bruno Stoufflet, Chief Technology Officer at Dassault Aviation and a recognized CSE expert, brought that message home with an insider’s detailed account of the field’s expanding role to attendees of the 2017 SIAM Conference on Computational Science and Engineering (CSE17), held this February in Atlanta, Ga.
Dassault Aviation designs and manufactures business jets, military fighters, and unmanned aerial vehicles (commonly called drones). It develops most of its simulation codes in-house, with assistance from external collaborators. Embodying the firm’s wide scientific awareness, Stoufflet cited the value of the finite element work of Thomas Hughes, who was in the audience and had just received the SIAM/ACM Prize in Computational Science and Engineering.
At Dassault, CSE plays a role in three phases of an aircraft’s life cycle: design, development, and post-delivery support. Design commands by far the largest share of Dassault’s CSE efforts, most of which promote a traditional engineering perspective (although newer stochastic approaches are emerging). Stoufflet’s long list of design-centered challenges included industrial-scale computational fluid dynamics (CFD), automatic shape optimization, multi-physics analyses, computational electrodynamics, surrogate modeling, uncertainty quantification for robust design, and the movement of all these computational tasks to exascale environments.
A typical computational task during the development phase is evaluation of the probability of rare events, such as a collision during the release of a store from an aircraft. While some of the variables are under the pilot’s control, others are not. Brute-force estimates are not feasible, so importance sampling is the primary tool. In a somewhat different direction, computation’s role in support following delivery typically involves data analytics that add value by justifying reliability estimates or guiding predictive maintenance, for example.
Stoufflet emphasized the multidisciplinary character of the design loop, which consumes the lion’s share of Dassault’s CSE efforts (see Figure 1). The design loop begins with every available option for architecture and technology on the table. Analysis and optimization in the design phase are driven by such specific disciplines as aerodynamics, structures, acoustics, propulsion, and aircraft systems. Formulating parametric models enables engineers’ exploration of the design space and their evaluation of sensitivities and risks. Each lap of the design loop ends with considerations of market and regulatory requirements like range, comfort, environmental restrictions, and cost.
Figure 1. Dassault Aviation’s multidisciplinary design loop. Image credit: Dassault Aviation.
Dassault’s bread-and-butter computational fluid dynamics tools have evolved from various finite element formulations with 10,000 nodes in the early and mid-1980s to the current operational CFD code that handles upwards of 20 million grid points in less than 15 minutes on 2048-core class machines (see Figure 2). Its steady Navier-Stokes solutions and unsteady eddy simulations are used during all stages of design. “We design for cruise conditions with the CFD code,” Stoufflet said. “A wind tunnel is used only for intermediate and final check-out.”
In the future, design teams hope to use their Reynolds-averaged Navier-Stokes steady solver to predict drag at cruising speed to within 0.5% accuracy in a typical compute time of 30 seconds. They also wish to improve models and better understand local flow physics to determine maximum lift and explore airframe acoustics more precisely. For example, predictions of noise generated by a landing gear bay are greatly influenced by the level of geometric detail included—local flow physics is at work—but a detailed computation on a 2048-core machine can require 15 days!
Figure 2. The evolution over 30 years of the capabilities of computational fluid dynamics for aircraft design. Image credit: Dassault Aviation.
Developing automatic shape optimization codes requires broad collaborations with the scientific community and “strong interaction with the design team to define and model the significant pieces,” Stoufflet said. The external collaborations have contributed automatic differentiation software, feasible direction sequential quadratic programming code, and a feasible arc interior point algorithm. These tools are integrated into a gradient-computation formulation that draws on partial differential equation control ideas of Jacques-Louis Lions, who also shares credit for inventing the parareal parallel-in-time algorithm.
When employed within the design loop, these tools have enabled a significant increase in the area of laminar flow on an aircraft wing near its fuselage, a reduction to nearly zero of an area of recirculation near an aircraft’s tail, an optimum balance between low-speed lift and high-speed drag in the shape of wingtip winglets, and optimization of separated flows in curved air ducts for unmanned aerial vehicles.
Dassault’s “total in-house control of tooling enables development of an optimization chain” that is in daily use, Stoufflet said. Shape optimization accelerates the early design cycle and offers wider options, while engineering analysis remains an essential step within the design loop.
Stoufflet’s list of sophisticated, engineering-focused computational challenges included a multi-physics computational challenge in aero-elasticity to correctly predict the influence of a missile carried on a fighter, varied approaches to computational electromagnetics involving complex materials and geometries, and surrogate models to facilitate interactive exploration of design alternatives.
As Chief Technology Officer, Stoufflet also faces a broader, computationally-based challenge in Dassault’s approach to the design process; the trade-off is robust design versus reliability-based design, or seeking to “manage uncertainties instead of adding margins.” He suggests that a designer “needs a new mindset” to think in terms of the “probability of not reaching a (design) objective,” rather than adding some arbitrary margins to a minimum drag requirement — for example, in hopes of accounting for the effects of possible twisting of the wing, changes in trailing edge camber, and the like.
As mindsets change, designers will need ways to propagate uncertainty through a system and access to computed response surfaces that incorporate second-derivative data, which in turn require fast CFD solvers and automatic differentiation tools.
At Dassault, Stoufflet anticipates “an unceasing effort to increase the efficiency of the design process,” continued exploitation of the benefits of quantifying uncertainty, increased confidence of engineers in stochastic approaches, and the exploration of “applications of data analytics that bring added value, e.g., justifying the correctness and reliability of machine learning” algorithms.
Addressing his audience directly, Stoufflet re-emphasized the importance of the CSE community’s rigorous approach while acknowledging that industry is more ad hoc. He requested that his listeners guide their own work by asking themselves, “What can we put into our framework that is rigorous and helps industry?”
Stoufflet’s CSE17 presentation is available from SIAM either as slides with synchronized audio, or as a PDF of slides only.
Read more about the latest developments in parareal algorithms in a report of another invited lecture from CSE17.