51³Ô¹ÏÍø

Researchers at 51³Ô¹ÏÍø have identified a hidden risk in AI-assisted engineering design. Their study reveals that models can appear reliable individually and when combined, yet still steer optimisation processes away from the best solutions, causing promising innovations to be overlooked. The findings have implications for aerospace, net-zero technologies and other complex systems, highlighting the need for multifidelity and multi-source design approaches.

Researchers from 51³Ô¹ÏÍø have identified a previously unrecognised phenomenon that could cause engineers to overlook some of the most promising designs for future spacecraft and aircraft.  

The team discovered that engineering models can appear accurate and reliable when assessed individually, and even when their interactions with other parts of the system appear well- represented, yet still steer designers away from the best solutions when used together as part of a complex system. They describe this phenomenon as the "Modelling Adequacy Paradox".  

Published in the study raises important questions about how next-generation aerospace systems are designed, particularly as engineers increasingly rely on artificial intelligence and automated optimisation tools to evaluate thousands of potential concepts.  

The research originated in Professor Laura Mainini's group in 51³Ô¹ÏÍø's Department of Aeronautics and was sparked by unexpected findings from computational experiments carried out by MEng student Livia Trambaiolo and Research Associate Dr Francesco Di Fiore on re-entry space vehicle design.  

When researchers compared conventional fixed-fidelity modelling approaches with more advanced multifidelity methods, they found something unexpected: the best performing design lay in a region of the design space that conventional fidelity models had consistently dismissed and never explored. 

"What surprised us was not that simplified models can be less accurate, which is already well known, but that models can appear fit for purpose for an individual discipline and even when coupled with others, while still steering the design search away from better solutions”, said Professor Laura Mainini. “The interactions may be captured and the predictions may look reasonable, yet promising regions of the design space can still remain hidden. The paradox is that a model can simultaneously provide valuable guidance and misdirect the automated design search, while the misleading effects remain largely invisible to designers.” 

Why the research matters 

Early-stage design processes rely heavily on simplified models to reduce computational cost and screen large numbers of possible concepts before more detailed analyses are introduced. The researchers found that even simplified models that appear adequate can reshape the design space explored by the optimisation process, causing promising technologies to appear unattractive or even infeasible.  

As aerospace systems become increasingly complex and derive more of their performance from interactions between physical technologies, software, control systems and AI, the consequences could become more significant. The researchers warn that AI-assisted and automated optimisation tools may be particularly vulnerable because they can rapidly explore design spaces using models that appear reliable while unknowingly reinforcing hidden biases in the search. As a result, potentially valuable innovations could be screened out before they are ever examined using higher-fidelity methods. 

The implications extend beyond aerospace. The researchers believe the same effect could influence the design of any complex engineering system whose performance emerges from interactions between multiple components, particularly as physical and digital technologies become increasingly interconnected.  

The paradox is that a model can simultaneously provide valuable guidance and misdirect the automated design search, while the misleading effects remain largely invisible to designers Professor Laura Mainini Chair in Aerospace Computational Design

The findings also raise broader questions about the growing use of AI-assisted engineering design. While automated optimisation and machine learning tools can accelerate innovation and help engineers explore thousands of design options, the researchers argue that practitioners need a stronger understanding of how modelling assumptions influence the results generated by these systems. Models that appear successful and trustworthy can still distort the design landscape and steer optimisation towards less promising solutions, particularly when multiple disciplinary models are integrated to represent complex engineering systems. 

A new approach to engineering design 

The team argues that future design processes should move beyond the traditional staged use of model fidelity, where simplified models are used to screen concepts early and higher-fidelity models are introduced only later in development. This becomes particularly important as AI-assisted optimisation is increasingly used to explore and screen thousands of possible designs automatically: if an apparently reliable model provides a misleading view of the design space, automation can reinforce that bias at scale and exclude promising concepts before more detailed evidence is introduced. Instead, the researchers advocate multifidelity and multi-source optimisation approaches that combine evidence from models of different types and levels of fidelity throughout the design process. By bringing higher-fidelity information into the search selectively and at critical decision points, engineers can identify when apparently adequate models are obscuring important opportunities and reduce the risk of automated optimisation converging on suboptimal solutions. 

Such approaches could help reduce costly redesigns, shorten development cycles and identify safety-critical issues earlier in the design process by introducing higher-fidelity evidence at critical points in the search without incurring the cost of using high-fidelity models everywhere.  

The implications are especially important for transformative challenges such as net-zero aviation, where progress may depend on unconventional aircraft configurations and new interactions between propulsion, structures, energy systems, software and control. If optimisation processes are guided by models that favour established technological assumptions, breakthrough concepts could be overlooked despite appearing viable when assessed using more comprehensive modelling approaches. 

Professor Mainini said: "As engineers increasingly adopt AI-assisted design tools, it is essential that we understand not only whether a model is accurate, but also how it influences the path taken through a design search. Multifidelity and multi-source approaches provide a way to challenge those hidden biases and support the development of safer, more efficient and more sustainable technologies." 

The researchers will now continue developing multifidelity and multi-source optimisation methods to support the design of safer and more sustainable aerospace technologies.  

The study was conducted by Livia Trambaiolo, Dr Francesco Di Fiore and Professor Laura Mainini from 51³Ô¹ÏÍø's Department of Aeronautics.