Applying Machine Learning To Accelerate TCAD Calibration

TCAD models are the fundamental building blocks for the semiconductor industry. Whether it is a new process node or a new multi-billion dollar fab, accurate TCAD models must be developed and calibrated before they can be deployed in technology development. While TCAD models have been around for (many) decades, their complexity is growing exponentially, as is the demands placed on the R&D engineers who are responsible for the models. Calibrating TCAD models continues to be a labor intensive and time-consuming effort.

During technology development, TCAD models play a crucial role in both pre-wafer and post-wafer phases. In the Pre-Wafer Phase, TCAD models are employed to predict trends based on physical laws, helping engineers explore and down-select design options which will be validated on engineering wafers.

In the Post-Wafer Phase, an accurate TCAD calibration is essential to align models with specific processes, ensuring reliability in further development stages. However, once initial wafers are fabricated, classical calibration methods become time-consuming.

Deploying machine learning (ML) in TCAD calibration

Calibrating TCAD models require many rounds of experimental wafers and hundreds of optimization loops done via Design of Experiments (DoE). These tasks are all done serially, meaning this is a very time and resource intensive effort. What if we could apply machine learning (ML) to the iterative parts of this process? An ML-assisted process could use simulations to run DoE’s in previously unexplored regions, as opposed to costly time-consuming experiments.

Consider the example shown below in creating a model using a manual method of calibration compared to an ML-assisted method. In both cases we input our parameters, run a sensitivity analysis, and input a second set of parameters. (In both cases, we would start with a set of input parameters, then perform a sensitivity analysis to identify the most impactful parameters.) In the manual method we would then run hundreds of optimization loops which can take several hours. This turns into hundreds of hours (1-2 weeks) of analysis to obtain the calibrated TCAD model. In the ML-assisted method, we have extra steps to run simulated DoE and train the ML. This could take up to a day. However, once we have our trained ML model the optimization loops take only milliseconds, so we can run thousands of loops almost instantaneously. The ML-assisted calibration method saves an enormous amount of time and resources.

Enabling AI and ML in the fab

In 2023, Synopsys rolled out a number of AI initiatives for silicon design, test, and fabrication. Synopsys Fab.da utilizes AI to optimize process control in a high-volume semiconductor fab. Synopsys is also bringing ML to the TCAD calibration process.

Sentaurus Calibration Workbench (SCW) utilizes ML to significantly enhance the value of TCAD models by enabling rapid and systematic calibration. ML-assisted calibration improves Quality of Results (QoR) while reducing calibration turnaround time. SCW provides an automated workflow that comes with pre-defined calibration modules and can easily integrate customer-created ML models.

Conclusion

The economics of semiconductor manufacturing continues to grow in challenges. There is always a race to the new process node in multi-billion dollar fabs, meaning every step in the process is critical to beating the competition. Calibrated TCAD models have always been a key to enabling a fab’s success. While calibrating these models has historically been a time-consuming step, we can now apply ML methods to optimize and speed up the creation of TCAD models. This will help in faster process ramps and greater ROI for new fabs.