An AI-enhanced multidisciplinary optimization framework for net-zero transport aircraft

aiaa.org . . An AI-enhanced multidisciplinary optimization framework for net-zero transport aircraft Andrés Mateo-Gabín, Marta A. Martín, Jose Fernández Montes, Sven A. Lanzan Ferran, Simone Mancini and Tim Klaproth Airbus Defence and Space SAU AIAA Aviation, July 21–25, 2025 Copyright © by Airbus Defence and Space SAU. Published by the American Institute of Aeronautics and Astronautics, Inc., with permission. . . Outline •HERWINGT project •MDAO framework •Generative AI for optimization •Conclusions 2 . . HERWINGT Project and scope 3 . . HERWINGT The Hybrid Electric Regional Wing Integration Novel Green Technologies (HERWINGT) project is one of the gamechangers to drive the transformation toward the decarbonization of aviation systems. It aims to design a novel wing, ideal for the future Hybrid-Electric Aircraft of the Regional Segment, and to develop architectures, structures, and technologies allowing higher integration of electric systems. Through these disruptive solutions, it will be pursued a 50% fuel burn reduction,at the aircraft level compared to a2020 state-of-the-art (SOTA) aircraft, in three different ways: •Pioneering wing configurations and improved aerodynamics leading to drag reduction and enabling a fuel burn reduction of 15%at the wing component level, compared to a2020 SOTA wing •Wing structures, more integrated systems, and new material technologies resulted in a weight reduction of 20%at the component level, compared to a2020 SOTA wing •The development of technologies enabling the wing for a hybrid-electrical use case: H2/Batteries + fuel systems using Sustainable Aviation Fuels 4 4 HERWINGT project web site: https://herwingt-project.eu/home. . . MDAO framework Development and industrialization 5 . . MDAO framework Environment •Gap between academic and industrial applications of MDAO is still large •High number of disciplines •Many in-house SW and internal know-how •Large set of requirements and constraints •Academic cases focus on novelty, sometimes overlooking the complexity of real systems such as aircraft 6 6 Our approach •Intuitive definition of systems and physical models (inspired on Entity-Component-System patterns and CPACS) •Decoupled definition of disciplines and solver implementations to add flexibility •Consistent definition of design and coupling variables •Automatic coupling of disciplines and generation of dependency graphs M. Alder, E. Moerland, J. Jepsen and B. Nagel, “Recent Advances in Establishing a Common Language for Aircraft Design with CPACS,” in Aerospace Europe Conference, Bordeaux, France, 2020. https://gemseo.readthedocs.io/en/stable/ . . MDAO framework Components •Containers of properties •Can be nested •Parent and Connections allow multiple hierarchies: geometry, fuel or electric systems… 7 7 Properties Previous segment Properties Previous segment Properties Previous segment Propeller Blade Properties Parent Connections Properties Section Section Mission •Concatenation of segments •Singly-linked list →missions can be branched . . MDAO framework Variable •Has a name, value and bounds Discipline •Physics simulation •Initialized from components and solver options •Defines its interface dynamically 8 8 Discipline Component Component Properties Variables Components Mission segment Properties Variables Previous segment Component Grammars Discipline information Variables . . MDAO framework. HERWINGT wing 9 9 . . Generative AI for optimization Best individuals 16 16 Design / Objective #1 #2 #3 Root chord 1.770 1.997 1.610 Kink position 0.507 0.725 0.790 Zero -sweep line 0.337 0.666 0.136 Tip/root chord ratio 0.719 0.924 0.640 Tip twist -2.353 0.450 0.559 Wing semi span 22.370 18.993 28.340 Wing drag 2179.569 2073.174 1943.249 Wing mass 2328.450 2427.229 3345.206 Fuel mass 15189.450 16454.15 14632.26 Fuel fill ratio 0.466 0.443 0.415 Design #1 Design #2 Design #3 . . Generative AI for optimization •Denoising Diffusion Probabilistic Models are widely used and show state-of-the-art performance •Starting from a completely noisy sample, they are trained to remove the noise incrementally, returning meaningful results 17 17 J. Ho, A. Jain and P. Abbeel, “Denoising Diffusion Probabilistic Models,” in Advances in Neural Information Processing Systems, Vancouver, 2020. 𝑝𝜃Ԧ𝑥𝑡−1 Ԧ𝑥𝑡= 𝒩( Ԧ𝑥𝑡; Ԧ𝜇𝜃Ԧ𝑥𝑡, Ԧ𝑦𝑡, 𝑡 , Σ𝑡Ԧ 𝐼) 𝑞𝜃Ԧ𝑥𝑡Ԧ𝑥𝑡−1 = 𝒩( Ԧ𝑥𝑡−1; 1 − 𝛽𝑡Ԧ𝑥𝑡−1, 𝛽𝑡Ԧ 𝐼) Σ𝑡=1 − ത𝛼𝑡−1 1 − ത𝛼𝑡 𝛽𝑡 •In our case, the design variables play the role of the image, and the objectives and constraints condition the model . . Generative AI for optimization Model architecture •The input vector contains the design variables, the time step and the conditioning •The output vector only contains the mean of the Gaussian noise to be added to the design variables •MLP with four layers •11 →512 + tanh •512 →1024 + ReLu •1024 →256 + ReLu •256 →6 •10% dropout during training 18 18 . . Generative AI for optimization 19 19 . . Generative AI for optimization Proposed population •102 new individuals •80 valid (22 were out of bounds) •Wing mass is well predicted •Fuel mass is somewhat less consistent 20 20 Accuracy Design space exploration Population size and # of populations . . Conclusions 21 . . Conclusions •Holistic approaches are essential in the design of present and future aircraft •MDAO methodologies can bring many benefits to industry, but they must be flexible enough to incorporate the existing know-how •In our framework, we have successfully abstracted the definition of the system from the MDAO loops, automating the connection between the two •The coupling with GEMSEO allows a wider range of applications, including uncertainty quantification or surrogate modelling •The framework has been tested in the HERWINGT use case, including a representative subset of common disciplines in aircraft design •Generative models based on AI could present an alternative to classical optimization algorithms •Deep architectures can capture highly non-linear behaviors, providing samples that match well the requested performance 22 22 . . The project is supported by the Clean Aviation Joint Undertaking and its members. 23 23 Clean Aviation is the EU’s leading research and innovation program for transforming aviation towards a sustainable and climate neutral future. As a European public-private partnership, Clean Aviation pushes aeronautical science beyond the limits of imagination by creating new technologies that will significantly reduce aviation's impact on the planet, enabling future generations to enjoy the social and economic benefits of air travel far into the future. Visit the website to find out more about Clean Aviation: www.clean-aviation.eu. Funded by the European Union, GA No 101102010. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or Clean Aviation Joint Undertaking. Neither the European Union nor Clean Aviation JU can be held responsible for them. aiaa.org Thank You . .