Access Official Technical PDF

Share Summary:LinkedInEmail

TCAD machine learning defect prediction offers a way to address a persistent challenge in advanced semiconductor failure analysis: locating defects that are difficult to observe directly. As transistor structures become smaller and more complex, physical inspection can be time-consuming, while electrical measurements alone may not clearly identify where a defect sits. A workflow that combines technology computer-aided design (TCAD) simulation with supervised machine learning can use those measurements to guide the investigation.

A study of bridging defects in a single-fin FinFET demonstrates this approach. Researchers generated simulated current-voltage (I-V) responses for defect configurations, extracted electrical features, and trained a Random Forest model to classify the defect location. The results illustrate how simulation-based data can support failure analysis when real failing devices are scarce.

Why FinFET failure analysis needs new approaches

FinFETs help manage short-channel effects that become more significant as conventional CMOS transistors are scaled. Their three-dimensional, non-planar structures, however, also make fault identification more challenging. Nanoscale non-visual defects can be difficult to locate, and the devices’ multilayer construction can make transmission electron microscopy (TEM) analysis complex and time-consuming to prepare.

Electrical nanoprobing is therefore an important part of defect investigation. But interpreting the measurements is not always straightforward: current can be affected by interactions among the defect, transistor and local interconnect. An unusual I-V response may indicate a problem without identifying the physical region responsible for it.

That distinction matters in failure analysis. A more focused estimate of defect location can help direct follow-up investigation, although the prediction is a guide rather than a replacement for physical analysis.

How TCAD machine learning defect prediction works

The workflow begins with a calibrated, three-dimensional TCAD model of a single-fin FinFET and its local interconnect. Researchers introduce modeled bridging defects at different positions and simulate the resulting electrical behavior. The study’s defect material was titanium nitride, with fixed dimensions of 5 nm, 18 nm and 3 nm along the modeled X, Y and Z dimensions, respectively.

Instead of relying on a large collection of naturally occurring failures, the method uses these simulated device configurations to build a labeled dataset. Each sample links an electrical response to a known defect region. In this case, the modeled area was divided into 10 regions, and the dataset contained 273 samples.

Measurable attributes were extracted from the simulated I-V curves and used as input features for a supervised Random Forest classifier. The model was evaluated using randomly split training and testing data, with cross-validation and grid search used to select model parameters. Because the number of samples varied among regions, class weighting was set to balance the classes during training.

What I-V characteristics reveal about defect location

The simulated responses show why multiple electrical features can help distinguish defect configurations. A bridge between the drain and channel within the fin increased leakage current compared with a control device. The study’s band-diagram analysis associated this behavior with a reduction in the source barrier, allowing more leakage under the examined off-state condition.

A defect between the source and channel produced a different response: the defective device showed a lower current level than the control. In the simulated analysis, a Schottky barrier between the defect and channel increased the source barrier, limiting the carriers able to pass through.

Bridging paths outside the fin also produced recognizable patterns. A resistive path between the drain and gate led to a linear relationship between gate voltage and drain and gate currents, while the source current retained a transistor-like characteristic. For a source-to-gate path, source and gate currents increased linearly with gate voltage, while the drain current followed the control device’s transistor-like behavior.

These differences provide a physical basis for classification. The model did not treat the I-V curve as an unexplained pattern: features describing its electrical behavior helped separate defect regions.

Model performance and influential features

Across 1,000 random training-and-testing splits, the Random Forest model using all features achieved an average accuracy score of 0.9612. The confusion-matrix analysis found that samples from all regions except region 5 were classified correctly in the reported evaluation, while also highlighting an imbalance in the number of samples for some regions.

Feature-importance analysis identified subthreshold-swing measures and threshold voltage as useful indicators. These parameters reflect how a defect can affect the device’s electrostatics, particularly when it is near the semiconductor region. Current-related features, including source-current slope and gate current, helped distinguish defects outside the fin, where a resistive path can produce increased gate leakage.

The researchers then retrained the model using features with importance scores above 0.05. This reduced-feature model achieved an average accuracy of 0.9629 across 1,000 runs, with reported accuracy of at least 0.86. In this dataset, selecting relevant features slightly improved average accuracy and reduced the influence of less useful inputs.

Practical considerations for engineering teams

The approach is most useful when simulation results represent the device behavior being investigated. The study reports qualitative agreement between simulated and actual I-V behavior for a gate-pattern defect, checked against existing failure-analysis cases. That provides a basis for the modeling approach, but the reported classifier results concern TCAD-generated defect configurations—not a broad validation across all real-world failures.

For engineering teams considering a similar workflow, several practices follow from the method:

  • Calibrate the device model: simulated electrical responses need to reflect relevant device behavior before they can support useful predictions.
  • Represent plausible defects: defect locations and configurations should be grounded in failure-analysis evidence where available.
  • Check class balance: uneven representation can affect classification and should be addressed during model training and evaluation.
  • Keep physical interpretation in view: feature importance is more actionable when connected to device mechanisms such as barrier changes or leakage paths.
  • Use predictions to guide, not replace, analysis: localization can help focus investigation, while confirmation still depends on appropriate failure-analysis methods.

Conclusion

Combining TCAD simulations with Random Forest classification provides a structured way to predict bridging-defect regions in a single-fin FinFET from electrical characteristics. The study’s high average accuracy on simulated data shows the potential of this method to guide failure analysis when physical defect samples are limited. For practical use, the central requirements are a representative calibrated model, carefully selected features, and evaluation that accounts for class imbalance. Together, these steps can make electrical data more useful for directing the next stage of semiconductor defect investigation.

Access Official Technical PDF

About The Author