Get the newsletter
Samsung · AI / ML · Filed Aug 15, 2025 · USPTO published Aug 13, 2026 · Aug 13 release

Samsung Patents AI That Checks Semiconductor Chip Quality Without Physical Testing

Every chip Samsung makes has to be measured for quality, but physical measurements take time and slow production. This patent describes a way to teach an AI to predict those measurements before they happen.

Normalization of semiconductor spectrum data using time-axis and wavelength-axis adjustments. Drawing from patent filing US 2026/0236840 A1.
Normalization of semiconductor spectrum data using time-axis and wavelength-axis adjustments.All 10 drawings ↓
Publication number
US 2026/0236840 A1(US20260236840A1)
Applicant
Samsung Electronics Co., Ltd.
Filed · Published
Aug 15, 2025 · Aug 13, 2026
Inventors
Junhyun NAM, Jongin LIM, Minsoo KANG, Minsu KO, Sungun PARK, Sungjoo SUH
Status when we published
Waiting for an examiner
Document
20 claims
US classification
706/12

How Samsung's virtual wafer measurement AI actually works

Making a computer chip involves hundreds of tightly controlled manufacturing steps, and after many of them, engineers physically measure the wafer to check if things went right. Those measurements are expensive, slow, and sometimes destructive. Samsung wants to replace many of them with an AI that predicts the outcome instead.

The system works by watching the data that flows through the manufacturing process and learning the patterns that connect process conditions to physical results. Over time, the AI gets good enough at spotting those patterns that it can tell you what the measurement would have been, without anyone picking up a probe.

For you as a consumer, the benefit is indirect but real: faster, cheaper chip production means fewer bottlenecks and potentially lower costs in the devices those chips end up inside.

From the filing · CLAIM 1
… training the second model, based on the features output from the first model, to generate virtual measurement data corresponding to the input data.

Translation: The system uses a second AI layer to predict the quality of the chip based on the patterns identified by the first layer.

How the two-model training pipeline generates predictions

The patent describes a training method for what Samsung calls a virtual measurement model, a two-part AI that learns to mimic the results of physical wafer inspections using only the data already generated during manufacturing.

The first component is a feature-extraction model, which takes raw process data (think temperature curves, gas flow rates, equipment sensor readings) and compresses it into a compact mathematical fingerprint. Crucially, it does this for both the new data point and a set of hand-picked representative data samples drawn from the full training dataset. Representative data here means examples that cover the range of conditions the process can produce, chosen so the model has useful anchors to reason from.

The second component is a prediction model that receives both fingerprints and outputs a virtual measurement value. By comparing the new fingerprint against known representative examples, it can estimate where the new reading falls. The training approach is similar in spirit to few-shot or example-conditioned learning (where a model uses reference examples to calibrate its predictions), applied to semiconductor metrology.

Key design choices include:

  • Training the two models together so the feature extractor learns representations that are useful for prediction, not just compression
  • Using representative data as explicit reference anchors rather than averaging them away
  • Keeping the pipeline processor-implemented, meaning it runs in software on existing compute hardware
From the filing · THE ABSTRACT
Provided is a method of training a virtual measurement model of a substrate. The virtual measurement model includes a first model and a second model.

Translation: Samsung is creating a two-part software system that learns how to inspect semiconductor wafers without needing physical tests.

What this means for Samsung's semiconductor yield rates

Semiconductor manufacturing is already heavily automated, but metrology (the science of measurement) remains a bottleneck. Physical measurements require specialized equipment, skilled operators, and time the production line does not always have. If Samsung can train a reliable virtual model, it could run quality checks continuously and in parallel with production, catching drift in the process far earlier than scheduled physical inspections would allow. That translates directly to better yield rates, which is the percentage of chips on a wafer that actually work.

The approach also fits a broader industry trend toward AI-driven process control, where models trained on historical production data gradually take over decisions that humans or simple rule-based systems once made. This particular filing sits at the narrow but high-stakes intersection of AI training methods and chip fabrication, an area well represented among the latest Big Tech patents from semiconductor and hardware companies investing in AI-assisted manufacturing.

Source: US 2026/0236840 A1 at the USPTO (PDF). Full patent text and figures are in the official publication.

Get our take in your Top Stories

Liked this breakdown? Add Patentlyze as a preferred source on Google, and our plain-English take shows up more often in your Top Stories the next time Samsung patent news breaks.

There are more where this came from

We read the new patents from Big Tech, carmakers and more and send you the ones worth knowing. Plain English, free, every week.

The drawings · 10 sheets

10 drawing sheets from US 2026/0236840 A1 · click any drawing to enlarge

Virtual measurement system overview, drawing sheet 1 from US 2026/0236840 A1
Sheet 1 Virtual measurement system overview
Training data generation process, drawing sheet 2 from US 2026/0236840 A1
Sheet 2 Training data generation process
Model training workflow steps, drawing sheet 3 from US 2026/0236840 A1
Sheet 3 Model training workflow steps
Group classification procedure, drawing sheet 4 from US 2026/0236840 A1
Sheet 4 Group classification procedure
Data criteria sorting table, drawing sheet 5 from US 2026/0236840 A1
Sheet 5 Data criteria sorting table
Spectral data normalization process, drawing sheet 6 from US 2026/0236840 A1
Sheet 6 Spectral data normalization process
Detailed model training steps, drawing sheet 7 from US 2026/0236840 A1
Sheet 7 Detailed model training steps
Reference data generation system, drawing sheet 8 from US 2026/0236840 A1
Sheet 8 Reference data generation system
Show the other 2 sheets
Samsung Patents AI That Checks Semiconductor Chip Quality Without Physical Testing patent drawing, sheet 9
Sheet 9
Samsung Patents AI That Checks Semiconductor Chip Quality Without Physical Testing patent drawing, sheet 10
Sheet 10
Patent filing page