AI Agents in Semiconductors Manufacturing
Challenges
Process engineers at TEL relied on manual, trial-and-error workflows to optimize semiconductor etching, causing delays, errors, and significant financial costs.
Solution
I designed and launched the GMA/SSA system, a team of AI agents that provided recommendations with reasoning, and a repository of validated configurations.
30% reduction in set up and etching time
AI as a partner within engineers' workflows
Well-documented results for scale
Company
Aitomatic
Position
Product and Design Lead
Year
2024
Market
Asia, B2B
Challenges
Imagine trying to bake a cake that looks and tastes perfect, but you only have a picture of the final product—no recipe, no ingredient list, and no instructions. Every failed attempt costs hours and money, leaving you frustrated and unsure if you’re making progress or just creating new problems.
This is the daily reality for process engineers at Tokyo Electron Limited (TEL) when optimizing semiconductor etching workflows. They must repeatedly adjust parameters like gas flow and temperature, analyze results, and iterate until they meet nanometer-level precision.
Each delay costs tens of thousands of dollars per hour, while inconsistencies across workflows create even bigger risks, like disrupted supply chains and wasted resources. One TEL engineer summarized it well: "I never know if my adjustments are solving the issue or creating new ones."
Solution
To solve this, I led the development of GMA/SSA, a system of AI agents. Think of it as a master chef guiding you through every step, adapting the recipe to your tools, and ensuring consistent results. GMA/SSA provides recommendations tailored to each machine, explains the reasoning behind its suggestions, and saves successful configurations for future use. It transforms a frustrating trial-and-error process into one of speed, precision, and confidence.
Through interviews with TEL engineers and analysis of historical workflows, I uncovered three core needs: engineers needed clear explanations to trust the AI’s recommendations, workflows had to adapt to the unique demands of different machines, and successful setups needed to be saved to reduce variability in future work.
Based on these insights, I and the team developed a system in which GMAs allowed users to define task-specific agents, tailoring workflows to their exact needs. AI-generated parameters came with reasoning and transparency, enabling engineers to validate and refine recommendations with confidence. The system also saved validated workflows, creating a living repository that reduced variability and sped up future setups.
During testing, TEL engineers emphasized the need for adaptability. In response, we added interactive visualizations and logic flows, ensuring the system was easy to use and adaptable to future needs.
Impact
Although GMA/SSA hasn’t been fully rolled out yet, early results are promising.
Simulations project a 30% reduction in processing times, which could save TEL around $20,000 per optimization cycle. Feedback from engineers has been positive, with around 92% expressing trust in the system and a readiness to use it. Testing also revealed a 15% reduction in variability across workflows, leading to more consistent wafer quality and fewer delays.
