Aerospace/defense manufacturers are understandably conservative when it comes to adopting generative AI. Whether airplanes, satellites, drones, naval ships, or army tanks, the equipment they produce is massively expensive and subject to strict military, government, and data security specifications.

Nonetheless, GenAI is making important inroads into this sector and will ultimately be used for a variety of productivity and innovation advances.

Here’s a roundup of the most common questions industrial manufacturers ask when it comes to where to aim their generative AI initiatives, where to focus future efforts, and the precautions and preparations to take.

Q: How is generative AI relevant for the aerospace/defense sector?

A: The typical GenAI use cases–such as creating text, audio or visual content, or using real-time chatbots to answer questions–may not seem immediately applicable for a sector that produces expensive, complex, highly customized and security-sensitive equipment. Commercial aircraft cost tens of millions of dollars. Satellites are at least $1 million a pop. Naval destroyers can run about $2.2 billion each, according to a recent report by the Congressional Research Service.

And all these machines must meet strict quality and reliability requirements, not to mention military specifications and government regulations and protections for proprietary or confidential data. Generative AI’s potential for hallucinations or sensitive data leaks might seem unacceptable to aerospace companies and defense contractors.

As it matures, however, GenAI will offer plenty of potential benefits, from aiding in equipment design to improving predictive maintenance and aftermarket service.

Q: Where should I get started?

A: The most popular area for GenAI today seems to be equipment design. Already, 41% of aerospace and defense organizations recently surveyed by Capgemini say they are piloting generative AI in 3D modeling to speed design, aerodynamically optimize parts, and reduce costs.

(It’s worth noting there is a nuanced difference between “generative design,” a method already in use, and “generative AI design,” which goes a step further by using deep neural networks and data from past designs.)

Look no further than NASA’s Goddard Space Flight Center, where Ryan McClelland, research engineer at the center, made news in 2023 by using commercially available generative AI to design specialized parts he calls “evolved structures.” In a fraction of the time it takes a human, this AI can design parts that weigh much less than traditional components. “You can perform the design, analysis and fabrication of a prototype part, and have it in hand in as little as one week,” McClelland says. NASA plans to use these evolved structures in future missions.

Boeing researchers are also exploring how generative AI models could optimize aircraft design, according to Todd Citron, Boeing’s chief technology officer. “[AI] can put more complexity into its electronic brain than a human can; it can optimize over a broader space,” he explained at a recent conference. “If you look at the machine learning-optimized structures, they just look inherently different. They almost look like an alien spaceship, because they don't have that regular structure that makes it simpler for one human to do a design.”

41% of aerospace and defense organizations recently surveyed by Capgemini say they are piloting generative AI in 3D modeling to speed design, aerodynamically optimize parts, and reduce costs.

Meanwhile, commercial aircraft manufacturers could use generative AI to more easily and effectively tailor design and manufacturing for specific operators, according to Raman Ram, leader of EY America’s aerospace and defense practice.

“If the aircraft is for an airline that is focused on total cost of ownership (TCO), I may want to include specific elements that maximize TCO,” he says. “But a different airline may want to optimize for cash-out and may care less about TCO.”

Q: What if we already use digital twins for product design?

A: Generative AI could reduce the time and cost of building digital twins used to simulate systems for product development and predictive maintenance.

“Traditional AI is still very manual, and a digital twin is an extremely manual process to build,” Rishi Ranjan, founder and CEO of software vendor GridRaster, told Aviation International News. Companies could eventually use large language models to remove a lot of that manual [coding] work.

“Large AI models … can look at text data and image data and start helping you create a digital twin for these automatically,” says Ranjan. He predicts that companies could ultimately build digital twins for a fraction of what it costs today.

More on Digital Twins

For more on digital twins, see “How Twins Are Driving the Future of Business” and “Digital Twins at Work: 8 Examples.”

Q: Where else should we anticipate applications of generative AI?

A: There are four other likely use cases of generative AI in aerospace/design:

Modernization of manufacturing processes and systems. Aerospace and defense manufacturing operations are hampered by the legacy software code still used in embedded systems, according to EY’s Ram. “This legacy code was developed by different suppliers, and has been bootstrapped over the years,” which makes integration difficult. Generative AI could serve as a translator.

“An interactive engine built on an underlying foundational model would be able to take the legacy code for a collision-avoidance module, for example, and revamp it into a modern programming language,” he says, which makes it easier to integrate.

GenAI also creates an opportunity for aerospace and defense manufacturers to connect production processes that are currently siloed by equipment line or plant.

By using GenAI to bridge the different technical landscapes, they could optimize organizational planning, manage exceptions, and react faster to situations that would otherwise result in a slowdown or a quality issue. It might also help predict such problems.

Quality assurance. GenAI could help quality control systems detect defects, irregularities, or variations during manufacturing. Algorithms can be trained to analyze images or data to identify problems, improving accuracy and reliability in defect detection, while also reducing time required for manual inspection. Once a problem is identified, generative AI could give technicians step-by-step instruction on how remedy it.

Maintenance and repair. Manufacturers have long used sensor data, analytics, and AI to predict when maintenance is needed or parts are about to fail, so they can head off time-consuming breakdowns. Generative AI can enhance this maintenance approach by automatically creating text or images that provide step-by-step instructions, including lists of required spare parts.

Such a system would allow maintenance staff to spend more time performing tasks instead of preparing instructions–enhancing productivity and reducing costs, according to Capgemini.

Also, aerospace maintenance and repair shops today can’t accurately estimate a turnaround time for an engine, says Ram, which creates frustrating delays for operators who need to get that craft back into service.

But by using historical data and generative AI, repair shops could give more accurate estimates. A chatbot, for example, could tell technicians what to expect, so the repair shop can provide accurate estimates and improve customer satisfaction. For example, “it could tell you that these six parts usually need to be replaced at this particular interval on this particular type of engine, as well as that it typically takes 73 days to get those parts,” says Ram.

Contract bidding and documentation. Government contracts require significant time and expense to prepare because they are highly detailed and customized. Contractors can use generative AI to automate at least parts of the process, creating first drafts based on templates, historical documents, or specific prompts from procurement officials, according to a report by Deloitte’s AI Institute.

The AI helps extract relevant clauses and requirements from previous bids, contracts and other documents, which human workers can then customize for the final contract. Similarly, generative AI could help draft technical documentation, saving time and money, according to Accenture.

Making satellite data more useful with GenAI

Although it’s outside the manufacturing realm. NASA, in partnership with IBM and Clark University, has developed a geospatial foundational model that will help organizations get more use out of the satellite data available to them.

As background, businesses use data from various satellites for many different purposes: tracking weather patterns, for example, or documenting environmental effects like tree-canopy loss in the Amazon.

Third-party satellite analytics companies such as LiveEO offer products that help businesses comply with the European Union’s Deforestation Regulation, which is intended to ban commodities linked with deforestation and forest degradation.

But it often takes a significant amount of time and effort to turn raw satellite data into something that a company can actually use.

Released last year, the model is trained on a year’s worth of data from a NASA satellite, and can be used to convert satellite data into customized, high-resolution maps of floods, fires, and other landscape features. The model is a significant advance because it will allow governments, businesses, and researchers to more quickly, easily, and inexpensively use satellite data.

“It enables businesses to save on the costs of developing training data sets for special applications,” says Manil Maskey, senior research scientist and project manager at NASA. “The hard part is already done for them. The model has already learned from the [foundational] data.”

Expected uses include estimating climate-related risks to crops, buildings, and other infrastructure; monitoring forests for carbon-offset programs; and developing predictive models to help design ways to adapt to climate change. The information can inform insurance and risk management decisions.

“Everyone is jumping on these foundational models because they are showing so much value,” says Clark University Professor Hamed Alemohammad, who was involved in the project. “If there is a wildfire, companies can use the satellite data to determine what was damaged, rather than sending someone out to validate that,” he says. It could also be used to determine weather conditions, which is important for aviation, transportation, and shipping industries, among others, he says.

Q: What should I do now to prepare for a GenAI initiative?

A: Get your data–and data governance–in order. Most aerospace/defense manufacturers are “still in infancy in terms of data quality and governance,” EY’s Ram explains. “We still lack good master data systems, and a single source of truth.”

Add to that the fact that these companies have a lot of proprietary data to protect, as do their suppliers. This results in a lot of skepticism about collaboration and data sharing, which results in a lack of transparency.

“We don’t even have visibility into the bill of materials beyond the first layer,” says Ram. “If you look at prime manufacturers, two-thirds of their costs are driven by their supply base, and a sizable portion of the supply base is made up of suppliers with their own intellectual property. The primes do not have visibility into that, and it just cascades down throughout the tiers.”

If you don’t have enough technical data, or if what you have is of poor quality or disorganized, generative AI can’t do much. If your component and subsystem suppliers don’t share their data, your overall aircraft design or digital twin will be incomplete.

A closely related issue is data security, which is extremely important in these industries. Defense contractors must meet certain regulatory requirements for security in order to even bid on government contracts. Given the reports of ChatGPT users inadvertently leaking proprietary data, this is a valid concern.

In general, regulatory requirements are leading more companies to zero trust, which requires technology that connects different data sources and manages access rights.

On the other hand, companies sometimes want to share data in order to improve models for the industry overall. Yet most do not have the governance in place to enable that. Aerospace companies and their suppliers, for example, have a common interest in developing new lightweight yet durable composite materials they can all use. But that data might be interspersed with proprietary data about their secret sauce. Without good governance of the models, companies cannot share data for use in large language models while protecting their IP.

When sourcing models from third-party providers, companies also need a way to allow such third parties to continually update their large language models (LLMs). Companies using third-party models will need to figure out how to share such information in a secure way.

Q: Are there other things that could slow down GenAI adoption?

A: The strict requirements for auditing and compliance will require some planning and oversight. Government contracts dictate a certain amount of data traceability and/or require audits of certain processes.

In a quality inspection system, for example, generative AI might help detect a problem, such as a component exceeding a specified tolerance, and even recommend potential fixes. But if traceability or an audit is required, companies need to make sure that a qualified human employee signs off.

Even in unaudited processes, a human should oversee generative AI to add context. If a chatbot provides instructions to a technician making a repair, for instance, it may include actions that are not possible. “Because a generative AI model lacks insight into the process that fixes the root cause and the people who are the operators, it might overlook potential limitations, such as infeasible movements or inaccessible spaces,” says a recent report from BCG. “As a result, a quality assessment is always required to ensure that the recommended remediation is practical.”

Q: What is the outlook for GenAI in this industry?

A: The complexity and cautious nature of aerospace and defense will make this sector slower than other industries to adopt generative AI. Meanwhile, they have data and integration challenges to overcome.

However, the ongoing knowledge worker shortage and intense focus on quality could accelerate GenAI adoption. For those that pursue GenAI initiatives, it could result in faster and better equipment design, more efficient manufacturing, and faster, less costly maintenance, all of which could improve the bottom line.