Automate And Speed Up TCAD Calibration With Expert Modules And ML Calibration Accelerator

Increasing complexity in semiconductor manufacturing has pushed the time to market and R&D costs significantly higher. In the world of AI, there is increased focus on efficiency to help address these issues simultaneously. Wafer-based learning, which is an iterative and linear process, is a key contributor to the increased semiconductor development time and cost. Technology computer-aided design (TCAD) has been used across the semiconductor industry to supplement wafer-based learning. TCAD entails the use of simulation to develop and optimize semiconductor process technologies and devices. TCAD models the effects of real silicon during the design phase so that fabricated devices will act as expected, and as a result, the engineers will need fewer silicon wafers for development. Beyond R&D and towards product ramp, TCAD accuracy to experimental data becomes critical for supplementing even larger sections of wafer-based learnings with simulations.

During early technology development, TCAD tools use physics-based models to predict device behavior for new architectures such as Si C-FET, vertical GaN FET, or 3D CMOS image sensors. The wafer processing during technology development happens in a window of process conditions, which is captured by specific physical TCAD model parameters. TCAD calibration is the process of tuning these physical model parameters so that the simulated results match measured data from real devices. Default model parameters often cannot capture process- and technology-specific variations to an extent that would lead to quantitatively accurate and predictive results. Proper calibration ensures that TCAD can be reliably used not only for qualitative device understanding but also for design optimization and technology development. Calibration using this real-world feedback of test wafers to fine-tune the TCAD parameters is an essential step to unlocking efficiency in modern day technology development.

Solutions for TCAD calibration

Synopsys provides comprehensive end-to-end TCAD calibration solutions via Sentaurus Calibration Workbench (SCW). SCW is part of the extensive TCAD suite featuring advanced process and device simulation tools, complemented by a robust graphical user interface (GUI) that streamlines simulation management and facilitates detailed analysis of simulation outcomes. In addition, Synopsys TCAD offers tools for interconnect modeling and extraction, providing critical parasitic information for optimizing chip performance.

This blog post explores two new developments in SCW that significantly accelerate the TCAD engineer’s calibration workflow. These include expert calibration modules that give users a 5X productivity boost by helping them to quickly start on the calibration workflow. In addition, a new companion product to SCW, Sentaurus ML Calibration Accelerator, reduces the time to calibration by greater than 5X. This post also discusses new ML-focused enhancements in SCW.

The value of TCAD increases with the accuracy of the modeling vis-à-vis hardware data. Initial models with a larger range of uncertainty are sufficient for the early stages of process development and technology pathfinding as fabs try out different device architectures. As the technology matures, and there is a greater volume of silicon test data available for the promising technology candidates, continuous TCAD calibration helps in bridging the gap of initial model predictions to the hardware data and can be used to refine the next processes and improve the device characteristics.

However, the process of TCAD calibration is not at all easy. Users often lack calibration expertise, leading to lower productivity and challenges in selecting the right parameters, calibration sequence, and parameter ranges. If the calibration through manual efforts takes too long, re-calibration of models to a test chip and subsequent wafer prediction may not happen fast enough. In that case, modeling adds little value, since process engineers cannot afford to wait for calibration.

The goal is to have automated calibration, rapid enough so that TCAD models can be updated using the measurements from each test. Then the calibrated TCAD models can be used to run a virtual design of experiments (DOE) and make an informed decision for the next wafer run. Artificial intelligence (AL) and machine learning (ML) can automate TCAD calibration and reduce the expertise needed by users.

Expert modules provide 5X productivity boost

SCW is the leading TCAD calibration tool that meets all the requirements outlined above. Unlike the manual approach, in which engineers execute one simulation at a time to calibrate models, SCW executes multiple parallel simulations upfront to create a TCAD-based ML model. The fast inference from this ML model is used to accurately calibrate the device in a very short time.

Setting up a calibration workflow requires expertise and time. Synopsys provides expert curated calibration modules for the user to quickly get started on their calibration workflow. The calibration modules include targeted use cases as well as broad range calibration scenarios (see Figure 1). They provide users a 5X productivity boost by helping them to start with 80% of the workflow pre-built. These modules must be customized for customer-specific needs, and the Synopsys team works closely with the users to achieve this. Once developed and customized, these models serve as an internal baseline that can be used for multiple technologies and shared among different teams for reuse. As shown in Figure 1, SCW comes with a specific set of targeted calibration modules as well as broader range modules such as CMOS and SIMS (secondary ion mass spectroscopy) calibration. Synopsys continues to develop new modules and work with users to help them customize the modules for their internal workflows.

Fig. 1: Expert calibration modules provide a 5X productivity boost to TCAD engineers for a quick start-up solution.

ML enhancements allow users to create their own modules and workflows

SCW continues to invest in ML capability, allowing the expert users to create their own modules. TCAD-relevant ML technology blocks can be reused quickly by TCAD engineers with limited ML expertise. The focus is on optimizing the DOE required to create ML models. Active learning leverages query by committee to interactively populate DOE and optimizes ML model accuracy. Active learning utilizes up to 50% less DOE compared to random sampling for similar accuracy. Two additional features of weighted and hypersphere sampling allow users to focus on high-value DOE and quickly remove outliers.

For the first time in TCAD, calibration with uncertainty-aware neural networks (UANNs) is used. These are ideally suited for the semiconductor world, where data may not be uniformly distributed in the parameter space. Among the many benefits of UANNs, they provide >5X faster training compared to Gaussian process networks. UANNs are also easy to set up as they do not go through the train-test-validate process of ML model development, helping to simplify the calibration workflow for TCAD engineers. In addition, they achieve similar accuracy as the parametric ML models while reducing DOE consumption between 15% (Sentaurus Device) to 70% (Sentaurus Topography).

Sentaurus ML Calibration Accelerator: Scaling compute for faster calibration

Even with ML-based approaches, calibration turnaround time remains a key bottleneck for engineers who want to reduce wafer dependency and R&D cost. To address this challenge, Sentaurus ML Calibration Accelerator has been introduced as a companion capability to SCW, designed to help reduce calibration turnaround time for ML-based workflows. The acceleration is delivered through an embedded, stackable licensing model, giving users a scalable knob to allocate additional compute for demanding use cases such as 3D calibration (see Figure 2). This ML accelerator delivers a >5X reduction in calibration times. It works out of the box with existing calibration workflows, is independent of the calibration use case, and is the key enabler to bring TCAD closer to product ramps in the semiconductor fabs.

Fig. 2: Sentaurus ML Calibration Accelerator is a companion product that helps user speed up calibration workflows.

In summary, automated calibration is essential to improve TCAD accuracy while minimizing expensive wafer costs. With ML-focused enhancements, expert modules, and calibration acceleration, SCW provides a one-stop shop for TCAD calibration needs. The recent updates provide a >5X productivity boost to create the calibration workflows and accelerate these workflows further by >5X to make TCAD calibration possible within days from getting new test data.