An article published by Semiconductor Engineering on August 27, 2026, calls for treating the introduction of artificial intelligence into radio-frequency, or RF, circuit design as a gradual path—not a project requiring replacement of the current platform or a complete rebuilding of the workflow. The idea starts from a practical problem: much of the capability of RF and microwave teams depends on accumulated institutional knowledge held by experienced engineers, including power-amplifier design methodologies, device characterization, and decisions made during the work that were not documented in a reusable form.
According to the article, losing this expertise lengthens the qualification period for new engineers, slows the evaluation of the design space, and may lead to a redesign that could have been avoided by testing a larger number of candidates. Therefore, the first question should not be whether the team will adopt AI, but how this capability can be built without disrupting the tools engineers rely on to produce actual designs.
Three Stages That Can Be Implemented in Parallel
The author, Daren McClearnon, whom the source identifies as the director of AI and machine learning products at Keysight, proposes a three-stage framework: capturing knowledge, coordinating tools, and accelerating exploration. This framework does not require completing one stage before beginning another.
Capturing Expertise and Turning It into Reusable Assets
The source notes that RF expertise is not usually found in code alone; instead, it is distributed across tools, equations, and practical experience in connecting simulation results with what appears in measured hardware. During the capture stage, this knowledge should be converted into forms that the team can run and share, and later hand over to software agents.
- Exporting schematics and layouts into Python code with adjustable parameters.
- Recording expert procedures as executable scripts.
- Converting the simulation flowchart into documented code that agent-based systems can understand.
The practical value of this step does not lie in using a language model by itself, but in making the internal methodology visible and repeatable rather than leaving it in the memory of specific individuals.
From Writing Scripts to Defining the Objective
The coordination stage builds on the knowledge that has been captured. Instead of having the engineer write and manage every script manually, the engineer can specify the desired outcome—such as designing an MMIC power amplifier with gain exceeding 20 dB and output power exceeding 28 dBm—and then leave it to a language model connected to the tools to find and invoke the appropriate tools.
According to the article, this does not mean moving directly to complete autonomy. Organizations currently often fall somewhere between manual programming and using early collaborative assistants that understand the engineer’s intent while human oversight remains close. The next step is specialized agents that handle multistage tasks with partial delegation, relying on the same coordination foundation.
Accelerating Exploration While Keeping Physics Under Supervision
The third stage focuses on increasing the number of designs that can be explored and reducing waiting time. The source mentions two main tools: surrogate modeling, which replaces computationally intensive electromagnetic simulations with fast neural-network-based approximations and may be two or three orders of magnitude faster for some structures; and AI-assisted optimization, which can handle a larger number of parameters simultaneously and find suitable solutions using fewer simulations.
However, acceleration has no practical value if the result is unreliable. The risks are particularly prominent in packaging, interconnects, coupling, power and ground integrity, and three-dimensional currents. These factors may make a fast and confident answer far removed from actual hardware behavior, subsequently making root-cause analysis of failures more difficult. Therefore, the physics used in modeling must be compared with actual measurements before expanding the scope of delegation to the system.
What Must Be Verified Before Scaling Up?
The framework also emphasizes the importance of tracking the source of the data, not merely the quality of the model. The team needs to know the source of the surrogate model’s training data, how it was labeled and cleaned, and what known assumptions and limitations it contains. This transparency is what defines the boundaries of speed and prevents it from becoming an invisible risk.
The source cites the example of Sphere Semi, an RFIC design company that faced the problem of exploring one design at a time. The company adopted a fully code-defined, Python-based flow to execute the generation, simulation, ranking, and optimization stages, while running hundreds to thousands of candidates through co-simulation of circuits and electromagnetics. According to the figures cited in the article, this approach achieved a 5- to 10-fold increase in productivity, a 6 dB improvement in isolation, and a 30% reduction in filter area compared with traditional manual design processes.
certi.news analysis: The actual change proposed by the article is not replacing the engineer with an independent agent, but transferring scattered knowledge into executable flows and then linking them to clear design objectives and measurable verification tools. This makes adoption less tied to a single platform decision and more tied to the quality of documentation, data, and measurements. However, the reported results concern a single example as presented by the source and do not by themselves prove that the same gains will be repeated in every design or environment. Questions such as the generalizability of surrogate models and their limitations when packaging, structures, or data change also remain in need of verification specific to each team.