Tech BusinessChips & Semiconductors
DRAM yield improvement explored with virtual twins and machine learning
Semiconductor Engineering describes a case study in which a calibrated virtual twin and machine learning predicted higher DRAM yield. The reported model result rose from 0.1% to 78.5% after process optimisation.

Key takeaways
- Semiconductor Engineering reports a virtual-twin case study for DRAM yield.
- The model included eight known systemic failure types.
- Predicted yield rose from 0.1% to 78.5% after machine-learning optimisation.
- The approach aims to reduce wafer experimentation during yield learning.
In this article
The report
According to Semiconductor Engineering, an article published on 19 November 2026 describes how process engineers can use virtual twins and machine learning to support DRAM yield work. The method uses a calibrated digital model of a manufacturing process to study failures and assess changes before more wafer experiments.
The article presents a case study, not independent test data. Its yield figures are predictions from the model, and the claims have not been independently verified.
It says the approach is aimed at faster root-cause identification. That matters because systemic defects can slow yield ramp and delay time to market.
The article positions virtual twins as a way to model variation inside a calibrated digital setting. It says machine learning can then evaluate several failure types and process changes at the same time.
How the virtual twin works

The article says engineers used a virtual manufacturing model called SEMulator3D. They paired it with a machine-learning tool named Fabtex Yield Optimizer. The aim was to predict process failures and reduce the need for physical wafer experiments.
A virtual twin is a digital copy of a process. In this case, the model was calibrated with silicon data so it could reproduce observed manufacturing behaviour. The article says this calibration helped the model reflect process variation more closely.
Virtual metrology measurements were set up throughout the process sequence. These simulated checks gave the model reference points for later failure analysis and optimisation.
The calibrated model was intended to support later failure analysis and process decisions. In other words, it was built not just to mimic the process, but to help engineers choose where to adjust it.
Finding multiple failure modes
The case study built a virtual twin for a DRAM memory device. It included eight known systemic failure types. These were failures linked to complex interactions among process modules, rather than simple random defects.
The examples included shorts, opens, and overlap defects in critical DRAM structures. The team defined virtual metrology checks, which are simulated measurements, for each failure. This meant the model could turn each failure into a measurable geometric condition and monitor it during the study.
By monitoring these conditions, the study could detect several failure mechanisms at the same time. The article says this happened within a single study, rather than separate experiments for each defect.
The article says this approach can reduce the need for wafer-based experimentation. Engineers can assess possible yield gains in the model before committing more physical wafers.
The reported yield result
The article says the model first estimated total yield at 0.1% under nominal, or standard, process conditions. After machine-learning optimisation, the predicted yield rose to 78.5%. Failure rates fell for all eight mechanisms.
The optimisation did not treat each failure separately. It searched for process targets that raised the overall pass ratio by evaluating several process changes at once. The article says the result was translated into inline fab targets, meaning settings that can be used during fabrication.
One highlighted example involved storage node landing pad open and short failures. The article says the optimised settings reduced all eight systemic failure modes in the model.
The article describes this as retargeting analysis. It says the exercise identified which process and structural settings need adjustment to reduce systemic failures further.
Why it matters
Semiconductor manufacturing is becoming more complex. The article says more process steps, narrower process windows, and new variation sources make acceptable yield harder to reach. Systemic failures are a key bottleneck during yield ramp, when a process is being improved toward target yield.
Conventional failure analysis can be slow and costly because it relies on physical wafer tests and statistical analysis. The reported approach aims to reduce that burden by letting engineers explore many process changes in a digital environment.
The article says the method can help engineers investigate a broad solution space with less silicon experimentation. If the predictions hold in production, it could shorten yield learning and time to market.
The article says the approach may help teams compare many options before choosing process settings. That could make yield learning more efficient during device development.
Sources
- How Virtual Twins And Machine Learning Optimize DRAM YieldSemiconductor Engineering
Spotted a mistake? See our corrections policy. Read how News by OnDevice works in our editorial policy.


