Create synthetic data with TCAD tools to train AI models that enable engineers to perform virtual fab runs.
The relentless pace of semiconductor development continues unabated. Despite the slowdown in Moore’s law, feature sizes continue to shrink as new geometries come online. Constant innovations in both fab processes and device design offer new opportunities but present new challenges. As in so many other areas of electronics, artificial intelligence (AI) is starting to play a significant role.
The complexity of modern chip fabrication makes experimentation in the fab very resource intensive. It takes too long and costs too much money for each turn of silicon to perform “what-if” experiments. The reliance on the empirical methods and wafers cannot be eliminated but it can be decreased if simulation tools are surgically deployed to help augment experiments. In between each learning cycle, engineers rely on technology computer-aided design (TCAD) tools to analyze the process and devices to predict results. The physics-based simulations provide high accuracy to experimental data, and TCAD engineers leverage them for “virtual fab runs” to supplement the wafer-based learnings.
Technology development (TD) engineers (process engineers, integration engineers, device engineers, test engineers, and yield engineers) typically do not have the modeling capability or expertise provided by TCAD tools. The silicon experiments provide them limited sampling, which makes it difficult to find an optimum process. This forces the TD engineers to rely on individual judgment and domain knowledge. TD engineers, as a result, are increasingly demanding modeling capabilities to help with their jobs. These engineers particularly want faster inference, while leveraging physics-based simulations, but lack expertise necessary to efficiently operate the TCAD tools.
Recent development of AI models fulfills the requirements of TD engineers for faster inference and ease of use. However, these models require a large amount of experimental data for training to be useful in a typical fabrication workflow. Especially in earlier stages of technology development, the amount of consistent data available in the fab is usually limited. Additionally, generating any data for AI model training requires fab resources that are getting more expensive with each node. The data is non-existent if a process engineer wants to understand the impact of the process change on electrical or mechanical properties of the semiconductor devices that have not been fabricated yet.
TCAD tools offer an effective alternative to create synthetic data that can be used to train AI models. The simulation data is significantly cheaper than experimental data. This lower cost allows the TCAD tools to cover the full parameter space needed to create high fidelity AI models that can be targeted at the TD engineer’s workflows in the fab.
Once trained, TCAD-based AI models allow TD engineers to simulate workflows. Figure 1 shows the TCAD-based workflow on top, where the experts run virtual experiments for “what-if” analysis. The AI-model based workflow on the bottom maps similar inputs to outputs, with AI models trained on TCAD synthetic data. The TD engineers are the target user for the AI models.
Fig. 1: TCAD and TCAD-based AI workflows.
Specifically, Al models enable engineers to perform virtual fab runs. The inference time for AI models is orders of magnitude faster than running TCAD simulations. With AI models, TD engineers get access to a tool that complements their wafer-based learning and helps them do their jobs. For the most part, they would just like to know how tweaking the process parameters will affect the performance of the fabricated devices. AI models enable them to get these answers in the shortest possible time without having to learn the TCAD tools.
One example of using the TCAD-based AI models is the design technology co-optimization (DTCO) workflow. A typical DTCO workflow entails connecting multiple TCAD tools with design tools and enabling closed loop optimization between technology and design. These workflows are not accessible to TD engineers due to the long simulation time and expertise required. The TCAD engineers can run parallel simulations of the DTCO flow upfront to collect synthetic data and generate the DTCO AI model for TD engineers. The immediate benefit to TD engineers is the ability to perform real-time co-optimization of the design and underlying technology. For example, the device engineer can quickly use the DTCO AI model to analyze the impact of device geometry designs on chip performance.
One additional benefit for foundries is that encrypted AI models provide an easier way to share their learnings with their customers without revealing key IP used in the process flow.
TCAD-based AI models also help process and integration engineers trying to refine the process and devices. One of their most frequent questions is how a change in a process parameter might affect device performance. Without AI models, TCAD experts must develop TCAD-based workflows to answer these questions for the process engineers. A simplest TCAD workflow might include process simulation followed by device simulation, with a relatively long TAT. Instead of sharing the TCAD workflow with the fab engineers, the synthetic data generated from these workflows can be used to train AI models. Process and integration engineers can understand the impact of changes in input parameters on the outputs of these AI models by tuning the “knobs” on the inputs.
In essence, AI models function as a “twin” for TD engineers’ workflows. Furthermore, multiple AI models can interact with each other to enable larger workflows. For example, one AI model might be trained to understand the impact of certain process changes on device characteristics, which in turn can feed into a DTCO AI model for design optimization.
As noted earlier, TCAD-based AI models are ideal when limited fab data is available. As the process starts to mature and the fab starts to generate consistent data, these AI models can further be enhanced by manufacturing data to create hybrid AI models with even better predictive accuracy.
The AI models must be considered in terms of the complete lifecycle of process development. TCAD’s physics-based synthetic data enables generation of AI models even before engineers run any silicon wafers. The AI models mature over time, consuming additional data and process knowledge, while increasing the value to the user from directionality to accuracy. These models eventually function as a store of knowledge for the fab, whenever there are requirements for that specific workflow in the future.
Fig. 2: Lifecycle of a TCAD-based AI model.
Synopsys provides tools to TCAD experts to manage the entire lifecycle of AI models through Sentaurus Calibration Workbench (SCW), as shown in figure 2. The AI model creation flow starts with identification of the user persona, such as device engineer for a DTCO workflow or process engineer for a process optimization workflow. The workflow is first captured by a TCAD expert in a TCAD deck. The experts use SCW to perform cost-effective calibration of the deck, which is then used to produce synthetic data. SCW provides users the platform to manage the simulation and manufacturing data. After choosing an underlying model architecture, the user can train, validate, and test the AI models. The models can further be updated if there is a change in the use case or new manufacturing data becomes available.
Every engineering team bringing up a new fab process, optimizing an existing one, designing new devices, or performing DTCO needs full-featured AI models for speed and accuracy. The ideal solution is available today here.
 |
 |
 |
 |
 |
 |
 |  |
Leave a Reply