Researchers Use AI to Design Better Analog Chips

Designing analog circuits is surprisingly hard. Unlike digital circuits, where signals are just 0s and 1s, analog circuits handle real-world signals like sound, temperature, and radio waves. Tiny changes in how transistors are arranged can completely change how the circuit behaves. Engineers often rely on trial and error, and even automated "genetic" algorithms that evolve designs can get stuck or need millions of tests.

Researchers at several companies and universities added a large language model (an AI that understands and generates text) to this design process. Their system, LLM-SPICEMixer, works like this: a genetic algorithm keeps a group of promising circuit designs. Instead of only making random mutations, an AI looks at the best designs and suggests entirely new ones. Each suggestion is then tested by a simulator called SPICE, which acts as the final judge. This combines the AI's creativity with reliable engineering.

The results were noticeably better. On a benchmark task, the AI-guided approach improved the final reward by 8.4 percent and the test reward by 8.8 percent compared with the genetic algorithm alone. For one test, the AI designed a transistor-level circuit that classified Iris flowers with 93.3 percent accuracy at normal operating conditions, and 85.9 percent accuracy across 17 different temperature, voltage, and process variations. The circuit itself is not useful, but it proves the method works.

The bigger picture: designing analog chips is a bottleneck in making everything from smartphones to electric cars and medical devices. If AI can help engineers explore better designs faster, it could mean cheaper production, better battery life, and more capable gadgets. This research is still early, and a human engineer still needs to check the results, but it shows one way AI can move beyond writing text and actually help build the physical world.