InsightFab
Knowledge Base/Virtual Metrology (VM): AI Prediction for Reduced Physical Metrology
Semiconductor Process6 min read

Virtual Metrology (VM): AI Prediction for Reduced Physical Metrology

This article addresses a critical challenge in semiconductor manufacturing: declining product yield due to slow and costly metrology, leading to delayed adjustments and inefficient production. The author shares a personal experience of relying on manual methods and enduring long waits for reports. The piece introduces Virtual Metrology (VM) as an AI prediction model that tackles these traditional metrology pain points, enabling proactive problem detection during the golden correction period and ensuring smoother production flow.

That Day the CPK Report Came Out, and Silence Fell for Three Seconds

I still remember a few years ago when the yield of a tool on our production line suddenly dropped. The client's Audit Team had just left, and then the Cpk report for the next batch came out, dropping directly to 1.08, with DPMO surging to 6210. All the managers in the Fab turned pale. The boss called a few of us engineers into the meeting room and simply said, "Find the problem; the production line cannot stop." At that time, we could only rely on rudimentary methods. Every batch was sent for measurement, and we waited for the report before adjusting tool parameters. The result was that each adjustment took several hours, severely dragging down production line efficiency.

Where the Problem Lies

To put it bluntly, one of the biggest pain points in semiconductor manufacturing is that "metrology is too slow and too costly." After each process step, you want to know if the results meet specifications, right? But a metrology tool worth tens to hundreds of millions cannot measure every single wafer. So, we can only rely on sampling metrology or wait until the product is fully manufactured to test final electrical properties. However, this misses the golden correction period.

Virtual Metrology (VM), in essence, is an AI prediction model. It doesn't actually measure; instead, it uses your existing "tool parameters" and "process data" to predict "metrology results." In other words, it transforms accumulated past process experience into a smart brain, telling you what the quality of the current wafer is likely to be. It predicts items that you "cannot measure" or results that are "not yet at the measurement point."

How It's Actually Done

It's not difficult, actually. Conceptually, it's a three-step process:

  1. Data Collection: You must first have a large amount of historical data. For example, for the past 1000 wafers, their Recipe parameters, chamber pressure, temperature, flow rates, etc., on tool A, as well as their final metrology results (e.g., film thickness, etch depth, CD value).
  2. Model Training: Treat these tool data as "inputs" and metrology results as "outputs," feeding them to the AI model for learning. Just like teaching a child to read, you show it many "combinations of tool parameters" and "corresponding results," and it will learn the relationship between them.
  3. Online Prediction: Once the model is trained, it can go online. When new wafers run through the tool, the model will instantly capture the tool parameters and immediately predict the metrology results for that wafer. It doesn't need to wait for the wafers to be sent to a physical metrology tool, saving a significant amount of time.

For example, we have an etch process where the most critical measurement is etch depth. Previously, we had to wait until hundreds of wafers were processed before sampling a few for measurement. Now, the VM model can predict the etch depth for every wafer. If the predicted value starts to deviate from the preset range, for example, gradually shifting from 1000 Å to 980 Å, we can detect it early, and even directly adjust the Recipe, instead of waiting for actual measurements to reveal that hundreds of wafers have already been lost.

The Most Common Pitfalls

To be honest, the biggest pitfall I've encountered is "data quality." At first, we thought it was simple: just throw all the data in. The resulting model's predictions were simply garbage. Only later did we realize:

  1. Unclean Data: Tool data often has issues like sensor abnormalities, data discontinuities, or simply missing records. Feeding such "dirty" data to the model will inevitably lead to predictable poor results.
  2. Poor Feature Engineering: We thought simply using the raw tool parameters would suffice. But often, the "rate of change," "cumulative values," or "combinations of different parameters" are the critical factors influencing the results. This requires process-savvy engineers to help define.
  3. Over-reliance: VM is, after all, a prediction, not 100% accurate. If you completely trust it and eliminate all physical measurements, the risk is too great. It should "assist" your judgment, helping you discover problems faster, rather than replacing all measurements. Therefore, physical measurements should still be performed regularly to calibrate your VM model.

One Thing You Can Do Today

Check if you have long-term accumulated historical data of "tool parameters" and "metrology results" available.

Want to try it yourself?

Every tool mentioned in this article is available on InsightFab — just upload a CSV to analyze.

Go to Tools

More in this category

Yield Engineering: From Die Yield to Wafer Yield

Ever wondered about wafer fab yield numbers but felt lost? This practical article introduces you to the world of wafer yield, using a vivid "pineapple field" analogy to clarify die yield, wafer yield, and what happens to defective chips. You'll learn how to pinpoint root causes from data, determining whether issues stem from equipment or processes, making crucial semiconductor knowledge accessible.

Yield Loss Classification: Random Defects vs. Systematic Defects

This article addresses the common factory challenge of unstable yield by differentiating between 'random defects' and 'systematic defects.' It provides a clear guide on how to identify the root cause of yield fluctuations, enabling more targeted problem-solving and significantly reducing time and cost spent on investigations.

Process Capability Assessment for Thin Film Deposition (PVD/CVD)

This article recounts a challenging experience where a newly promoted senior engineer faced intense pressure from management over a thin film deposition report showing a CPK of 1.08. It then provides an accessible introduction to thin film deposition, explaining common techniques like PVD and CVD. Readers will learn the critical importance of evaluating thin film process capability and the numerous intricacies involved in manufacturing these fundamental layers in chip production.