Electronics

SandBox Using AI, Hybrid Metrology to Cut Costs, Boost Yields

[ad_1]

//php echo do_shortcode(‘[responsivevoice_button voice=”US English Male” buttontext=”Listen to Post”]’) ?>

The hybrid metrology image processing tool Sandbox, which began selling last month, promises to improve metrology accuracy for etch and deposition steps, streamline experiments with new process recipes, and ultimately reduce process technology development costs. The tool, Weave, works with the SandBox Studio AI modeling platform to extract and analyze profiles from scanning electron microscope (SEM) and transmission electron microscope (TEM) data using machine learning (ML) algorithms, reducing manual measurement tasks for engineers.

Process tech growing more complex

Modern chip fab tech costs billions of dollars to research and develop, and it keeps getting more complex and expensive: Companies need to do research to discover the right materials for their needs, then design transistor structures that promise to meet power, performance, and area goals. They must then develop the right process recipe that takes into account how these materials and transistor structures are handled using actual fab tools.

Currently, the latter step in process development is carried out through trial and error, necessitating significant money (materials and tools are expensive) and time investments (it takes weeks or even months to process a wafer).

Meghali Chopra (Source: SandBox)

“Minimizing the time it takes to get to the right process recipe is really important for our customers,” SandBox Semiconductor CEO Meghali Chopra told EE Times. “We use our computational-modelling platform to help process engineers simulate process outcomes in order to get to their recipe faster.”Meghali Chopra (Source: SandBox)

One crucial thing to consider when developing a new process technology is to accurately understand device structures and how are they made: It’s particularly important for novel types of 3D NAND and DRAM memory, as well as gate-all-around transistors. An efficient way to accomplish this is to use metrology data gathered from multiple tools in a fab, since no single tool can address all challenges. That is called hybrid metrology.

“We believe that there is not going to be one single, beautiful metrology tool that is going to solve everybody’s problems; instead, we leverage a hybrid metrology approach,” Chopra said.

The problem is that getting data from multiple sources is prone to errors, time consuming and sometimes impossible. Process development engineers often dedicate nearly 20% of their time on manual measurement of scanning electron microscopy (SEM) and transmission electron microscopy (TEM) images, according to Sandbox. This is where Weave comes into play.

“We automate that extraction,” Chopra said. “So, we help process engineers save about 20% of their time, just by purely automating this [extraction] function.”

The tool aggregates data from different sources. It uses ML techniques to enhance the precision and accuracy of metrology by analyzing profiles from multiple sources, such as SEMs, TEMs, optical critical dimensions (OCD) spectroscopy and cross-sectional data. This lets engineers reduce manual efforts (by cutting down the number of experiments by up to eight times) and boost the accuracy of metrology—and that delivers the deeper knowledge they need to enhance yields, SandBox said.

SandBox Studio AI. (Source: SandBox)

By aggregating and rectifying data from diverse sources, SandBox streamlines it for analysis and integration into its modeling framework and enables engineers to use this data to make informed process predictions. Unlike generic AI models, SandBox uses semiconductor-specific behavioral physics models tailored for semiconductor production process development. Those models consider different materials and how actual fab tools work.

“So, using all that metrology data, engineers can now leverage all the tools they have at their disposal to actually make predictions across their process base,” Chopra said. “We are able to do this because we have reduced behavioral physics-based models, which enable process engineers to actually model their process space. We are running physics simulations, and then we use machine learning to augment those simulations. So, all the simulations and predictions we do are constrained within a physical framework, which tends to make our predictions more accurate.”

In addition, SandBox offers semiconductor-specific visualization tools and a tailored user interface.

“Then we have visualization tools that are specific to the semiconductor space,” Chopra said. “These tools that our users need in terms of digesting really large multi-dimensional process steps sets, we have a tailored UI experience.”

Working with the industry

Developing physics-based models and ML tools requires SandBox to work with the industry to understand how fab tools work, how materials work with these machines, and what chipmakers want to do. SandBox works with foundries, memory makers and manufacturers of etching and deposition tools, including Applied Materials and LAM Research. Meanwhile, the software doesn’t learn from data supplied by toolmakers.

“The software is learning purely from the empirical data that is being input into it,” Chopra said. “That is one of the key values of our approach is we can work with any type of edge or deposition tool or any type of metrology tool. The inputs for the software are essentially the geometry, which the process engineer can input himself, the OCD measurements, cross sectional SEM measurements, and then the experimental process parameters.

“So, no information from the tool is being fed into the software. With just the metrology data and the process parameter information, the AI model then learns from whatever has been provided to it. It learns from all the different types of metrology, and then it leverages our model libraries to figure out what is going on in the system and then we can make these predictions across the process space.”

Weave essentially helps engineers quickly correlate their observations with process parameters. The current focus of the software is on etch and deposition, which underlines importance of the platform for 3D NAND and DRAM makers, as well as foundries in the era of gate-all-around transistors. But the methodology can be applied to any process type, the company said.

Essentially, the software interprets metrology data in relation to process inputs, providing insights into the ongoing processes. This doesn’t necessitate expertise in a specific tool but rather a profound understanding of numerical modeling to accurately establish these connections.

“Because both tool makers and chipmakers do this type of process development, we work across the board with both chipmakers and toolmakers on a large range of applications, because process development is essentially the core of what they do,” Chopra said. “Every day most of these companies are just working on process development. So, either a process-yield problem, a critical dimension uniformity problem. These are a quarter of what they need to solve.”

The SandBox platform was designed to enable users to create their own computational models, eliminating the need to digest vast process spaces themselves. By using the AI platform, they can decrease the number of experiments and enhance process quality through a data-driven, quantitative strategy to define process parameters, which ultimately promises to reduce process development costs and maximize yields.

[ad_2]
Source link

Related Articles

Leave a Reply

Your email address will not be published. Required fields are marked *

Back to top button
Translate »