← sources

LLMs are already inside fabs, so why aren't they solving core yield and reliability problems yet?

source: reddit (reddit:semiconductors)published 2026-10-18 18:50ingested 2026-10-19 02:41

retained text

**(Quick note from OP: English is not my first language. As a non-native speaker, using AI to polish my draft already expresses my thoughts much better than the best English I could write on my own.**

**My only goal here is to genuinely discuss and exchange ideas with folks in this community about how AI fits into the semiconductor world. If the AI-polished wording bothers you, please feel free to just ignore this post. For anyone open to exchanging real ideas, I really appreciate your thoughts.)**



Following up on my last post here about AI in semiconductor manufacturing, I got a lot of great comments from folks in this sub about hallucinations, lot hold times, and how new tools fit into fabs that already have plenty of software. It made me reflect on where LLMs are actually being used in fabs today versus where they hit a wall.

If we are being fair, LLMs are already working inside many fabs. Most leading fabs have mature internal cloud platforms, plenty of compute, and local LLM deployments running safely inside their firewalls. Their data is also comprehensive and getting better every day as fabs constantly push equipment vendors to expose richer, higher-frequency tool data. Because of that, many engineers are already using internal LLMs every day to generate slides for morning and evening passdown meetings or to help draft basic scripts.

So at the peripheral level, LLMs are already useful. Returning to the question from the last thread though, fully replacing fab engineers is technically impossible today, and merely helping engineers make passdown slides faster is not the core value an enterprise fab really needs. The real needle-mover is solving hard problems on the line, especially deep yield excursions and reliability issues. And the bar here is getting much higher. As smart vehicles and autonomous driving scale up, automotive chips are becoming a massive driver of wafer demand. Automotive-grade requirements push yield consistency and latent reliability screening to a level where rough sampling or slow manual correlation simply cannot keep up.

Yet when you move from drafting passdown slides to investigating a real yield drop or a wafer-level reliability failure, a generic LLM quickly hits two and a half major roadblocks.

1. The first roadblock is that generic LLMs lack an execution mechanism for empirical scientific computing. Whether it is a defect team decomposing spatial signatures and filtering adders across wafer maps, a PIE or yield team running controlled commonality tests across a thousand process steps while ruling out scheduling weights and PM confounding factors, or a reliability team fitting Weibull tail distributions to separate early manufacturing defects from intrinsic wear-out, the core work is rigorous mathematical and statistical computation. In a fab, a number means nothing unless it is tied to explicit statistical scope, edge exclusion rules, reference group conditions, and reproducible calculations. A language model predicting tokens cannot do that on its own.
2. The second roadblock is folding past engineering experience into the actual data preparation and investigation pipeline. An off-the-shelf LLM knows textbook semiconductor physics, but it does not know how a specific fab prepares context across dozens of steps or how veteran engineers systematically rule out false leads. Simply dumping historical 8D reports or handover decks into a text search database does not bridge the gap between a past case and the live multi-step data preparation required for a new excursion.
3. The final half-roadblock, which is just as critical in practice, is user experience and engineer habit. Having worked in EDA and supported engineering teams before moving into AI, one thing I learned the hard way is how difficult it is to force engineers to adopt a brand-new habit. In a fab, individual calculation blocks are often already covered by existing tools. Engineers are used to a visual, evidence-first workflow where they inspect spatial maps, compare distributions, and verify raw records directly. Dropping a plain text chatbox in front of them goes completely against how they actually observe and verify problems.

That is why a standalone LLM is not the solution, but an agent architected specifically for this scenario is. Building an agent that actually works on the fab floor requires the designer to deeply understand four things at once: the boundaries of LLMs, the empirical scientific computing behind fab data, the established troubleshooting workflows in the fab, and the day-to-day habits of the engineers using the tools.

**To make the question concrete for those of you in yield, PIE, defect, or reliability: when a tricky excursion hits (say a post-PM edge ring signature or a wafer-edge parametric tail shift), where does your current workflow or tooling actually stall out? Is it stripping out upstream inherited defects, de-biasing high-throughput tool routing in commonality checks, or pulling the right multi-step context from past cases?**

document fetched 2026-10-19 00:31 · canonical https://reddit.com/r/Semiconductors/comments/1x9b1qf/llms_are_already_inside_fabs_so_why_arent_they