The AI design features unusual and efficient, circuitry patterns. Credit: Emir Ali Karahan, Princeton University. Engineers usually design wireless chips using familiar patterns and templates, refining them through simulations and painstaking testing. But a series of studies is proposing a different approach: a deep-learning-enabled design process for creating circuits and components. Using artificial intelligence (AI), researchers at Princeton University and IIT Madras demonstrated an “inverse design” method, where you start from the desired properties and then make the design based on that. The algorithm created some weird chips that seem to work really well.

But there’s a catch: no one really knows why they work so well. “Humans cannot understand them, but they can work better,” said Kaushik Sengupta, the lead researcher, a professor of electrical and computer engineering at Princeton. AI at the helm The AI-driven method focused on designing wireless chips, a type of computer chips used for high-frequency applications like 5G networks, radar systems, and advanced sensing technologies. These circuits power innovations in everything from radar systems to autonomous vehicles but their development is notoriously slow.

Engineers would start on predefined templates and manually optimize or improve designs through iterative simulations and testing. This method is time-consuming and challenging. It also requires a high degree of expertise, which limits just how much (and how fast) you can improve. This is where the new study comes in. Whereas the previous method was bottom-up, the new approach is top-down. It’s more efficient, because you create a system that does exactly what you want; but it’s much harder to get it to work. The researchers trained convolutional neural networks (CNNs) — a type of AI model — to understand the complex relationship between a circuit’s geometry and its electromagnetic behavior.

These models can predict how a proposed design will perform, often operating on a completely different type of design than what we’re used to. × Thank you! One more thing... Please check your inbox and confirm your subscription. The neural networks worked alongside optimization algorithms that searched for layouts meeting the engineers’ specifications. The researchers didn’t simply accept whatever the AI produced. They checked designs with conventional electromagnetic simulations, fabricated physical components, and measured their performance. The study showcased a range of use cases, from simple one-port antennas to complex multi-port RF (radio frequency) structures like filters, or power dividers.

The AI-designed compact antennas that function across two distinct frequencies, improving performance for multi-band devices. Within minutes, it synthesized filters with precise band-pass characteristics, a task that would have taken days or weeks before. The ability to rapidly design high-performance circuits could accelerate advancements in telecommunications, autonomous systems, and beyond. This approach empowers engineers to focus on innovation rather than routine optimization.

Yet perhaps the most exciting part is the new types of designs it came up with. Completely new concepts “We are coming up with structures that are complex and look randomly shaped, and when connected with circuits, they create previously unachievable performance,” says Sengupta. The designs were unintuitive and very different than those made by the human mind.

Yet, they frequently offered significant improvements. “Classical designs, carefully, put these circuits and electromagnetic elements together, piece by piece, so that the signal flows in the way we want it to flow in the chip. By changing those structures, we incorporate new properties,” Sengupta said.

“Before, we had a finite way of doing this, but now the options are much larger.” In August 2025, the Princeton group published another approach called Dall-EM.

This time, the researchers used diffusion models (the same broad family of AI techniques associated with image generators) to create electromagnetic structures that met specified performance targets. Instead of repeatedly testing and modifying candidate designs through an evolutionary search, the generative system could produce suitable structures much more quickly. The researchers reported that it reduced design time by at least a factor of ten compared with their earlier predictive AI approach, with some designs converging in seconds. Then came another advance.

At the February 2026 IEEE International Solid-State Circuits Conference, the team presented a design framework that brings together circuit architecture, component selection, electromagnetic design, and physical layout. Another June 2026 study addressed a different practical challenge: producing components that must fit within a prescribed physical area. AI-generated shapes may look impressive on a computer screen, but real chips have strict spatial constraints. These studies mark a pivotal moment in engineering, where AI not only accelerates innovation but also expands the boundaries of what’s possible.

However, researchers don’t see AI taking over chip engineering by themselves just yet. “There are pitfalls that still require human designers to correct,” Sengupta said. “The point is not to replace human designers with tools. The point is to enhance productivity with new tools.

The human mind is best utilized to create or invent new things, and the more mundane, utilitarian work can be offloaded to these tools.” A promising programme hits a classic obstacle The growing ambitions of AI-designed chips have also encountered a more conventional problem: funding. In May 2025, the National Semiconductor Technology Center announced a nearly $10 million award for a Princeton-led effort to develop AI-powered design tools for wireless chips.

The programme brought together researchers and industry partners, with Sengupta directing the research. But just a few months later, the project received a stop-work order. In August 2025, the US Department of Commerce withdrew support from Natcast, the organisation overseeing the programme, following a dispute over how the organisation had been established.

This was part of a wider set of decisions by the Trump administrations against programs started during the Biden administration. Sengupta’s team tried to seek out alternative funding, but struggled. An August 2026 government audit confirmed that Natcast had ceased operations in September 2025 and that the federal semiconductor research programme had undergone substantial cancellations. It’s unclear exactly what’s the status of Princeton’s original award. The setback hasn’t stopped the scientific work, as the group’s subsequent publications demonstrate.

But it illustrates how difficult it can be to turn an experimental advance into a widely usable technology, even when the underlying results are promising. Then, of course, there is the big-picture question. The black box How comfortable are we with these designs we don’t fully understand?

What happens if something goes wrong with this design? No doubt, AI will play an increasingly important role in how we design things, and chips are no exception. Yet for the most part, we still have no transparency on how AI arrives at its designs. This makes it difficult for engineers to fully understand or predict the behavior of these circuits under all conditions. This “black-box” nature could lead to unforeseen failures or vulnerabilities, particularly in critical applications like medical devices, autonomous vehicles, or communication systems. AI-generate designs aren’t necessarily inexplicable. Their behavior still follows the laws of electromagnetism, and engineers can test how they perform.

But it can be really hard to understand why a particular arrangement of metal traces produces a particular result. Engineers must establish how the devices respond to manufacturing variations, temperature changes, and operating conditions beyond those used in their initial tests.

The original research demonstrated measured performance in fabricated prototypes, but that doesn’t establish long-term reliability in commercial products. As Sengupta mentions, the best path forward is likely a partnership between humans and AI. AI can explore circuit layouts that human designers might never consider, while engineers remain responsible for deciding which designs deserve to be built, tested, and trusted. But there is a danger in becoming too reliant on systems whose decisions we don’t fully understand. A design that works brilliantly in the lab may behave differently under conditions nobody anticipated.

And if engineers lose the expertise to question, troubleshoot, or improve what AI produces, even small failures could become difficult to diagnose. AI can be a powerful tool for innovation, but its success should never become an excuse to surrender human understanding and oversight. Journal Reference: Emir Ali Karahan et al, Deep-learning enabled generalized inverse design of multi-port radio-frequency and sub-terahertz passives and integrated circuits, Nature Communications (2024). DOI: 10.1038/s41467-024-54178-1 Transparency note: This article was originally published in January 2025 and has been updated to include new research developments, additional context, and a more detailed discussion of the risks and limitations of AI-designed chips.

This update is part of our ongoing effort to revisit and improve previously published stories, ensuring our coverage remains accurate, relevant, and informed by the latest scientific evidence.