Making Chips Yield Faster At Leading-Edge Nodes

Simulation for semiconductor manufacturing is heating up, particularly at the most advanced nodes where data needs to be analyzed in the context of factors such as variation and defectivity rates.

Semiconductor Engineering sat down with David Fried, corporate vice president of computational products at Lam Research, to talk about what’s behind Lam’s recent acquisition of Esgee Technologies, which develops software for modeling and simulating plasmas and reactive flow systems used in a variety of industry segments. The discussion focuses on what’s changing in the manufacturing flow, and why simulating what’s happening during the etch and deposition processes has suddenly emerged as an essential path for reducing costs and time to yield. What follows are excerpts of that conversation.

SE: Why does the chip industry need plasma simulation, and why now?

Fried: There are lot of the challenges both in deposition and etch, that require us to push hard on specific plasma control capabilities. Much of that work is in hardware design — physically designing the chambers and the equipment — as well as process recipe design and control. So, it’s not only about how you build the chambers, but also about how you use various features of the chambers and the equipment to control the process behavior. If you look at some specific examples of challenging chamber activities, such as very high-aspect ratio etching used in advanced 3D NAND manufacturing, this plasma control is critical. Much of what we’re doing to drive the state-of-the-art in chamber processing relies on complex control of plasmas in our chambers. This extends from etch to deposition, as well, since plasma-enhanced chemical vapor deposition (PECVD) and plasma-enhanced atomic layer deposition (PEALD) tools are in common use across the chip industry.

SE: So basically, the process is getting more granular?

Fried: Yes, and as we increase the number of controls and number of knobs that can be controlled, you open this massive space of process options. If you experiment with those options from a hardware design perspective and from a wafer-based process development perspective, that becomes inordinately time-consuming and expensive. Simulation can dramatically cut this time and cost, both on the hardware side and on the process engineering side.

SE: Would this apply to mature nodes, or is this only for advanced nodes and packaging?

Fried: At mature nodes, if we can improve process performance or defectivity through plasma control, that’s a huge advantage for our customers. Simulation can be applied to assist in that kind of development. But the primary application will be in advanced technologies, the most advanced etch and deposition capabilities, such as those used in leading-edge logic, DRAM, and NAND. That’s where we’re designing the most advanced chambers.

SE: At the most advanced nodes, we’re hearing about logic with silent data errors due to manufacturing defects, and they don’t necessarily get caught because these chips are highly customized and produced in smaller batches. Will this have an impact?

Fried: The first order of business is to produce chips in an extremely uniform manner, regardless of whether that’s a data center chip or a mobile phone chip. Each of these chips is using the same process technology and the same processing equipment. So what we’re really talking about is quality and control of a particular process technology. Designs will vary, and certainly there is a lot of interesting work on the design side with design technology co-optimization. But process control requires the best process technology from a capability, variability control, and quality perspective. Having the ability to drill one channel hole in a 3D NAND array at the exact dimensional specifications is quite important, but we have to drill trillions of those holes on individual wafers with precisely the same dimensions and precisely the same quality. This requires precise chamber-scale control — controlling the conditions in a process chamber so they are very consistent for every wafer that goes into that chamber. That is what goes into designing the most advanced etch or deposition tools. The question then is, ‘How do we use chamber hardware options in process recipes and in our controls to maintain the required level of uniformity?’ Everyone in the value chain must do their part to reduce variability.

SE: If there’s some sort of irregularity in one of the gases being used in a chamber, and that’s coming out at a different rate than another chamber or it’s not as pure as you need, can you catch that and adjust for it?

Fried: There are two sides to that question. First, can we understand that effect? That’s where simulation has a huge advantage. Being able to simulate these variations using computer models and quickly review the results is where simulation really shines. It provides a depth of understanding of variation effects and the sensitivity of the process to those variations, which allows you to prevent those variations. There will be certain variations that have little to no effect on quality or yield, and you don’t care about those variations. In certain cases, if the gas varies it may not matter. You can prioritize your work to minimize variability based on simulation results, and then prevent issues that would affect quality or yield. It’s not so much about being able to catch these variations in real time. It’s more about having a depth of understanding of the sensitivity of those variations, preventing the ones that need to be prevented, and being able to have a process that has acceptable specifications.

SE: How do you digest all that data and apply it effectively?

Fried: We have quite a few sensors in our chambers, and we produce a large amount of data from those sensors. But having a depth of understanding about what all that data really means is critical, especially when you are discussing plasma physics. Some process effects are relatively easy to understand. They’re relatively linear, so there’s only one discipline needed to model the results. If I’m doing wet processing, I need to understand the chemistry of what’s going on. When I’m doing annealing, I must understand the thermal physics of an anneal. These are relatively straightforward disciplines. But with plasma processing, it’s a different story. There are chemistry, flow, particle physics, and electrical circuit effects. All these effects are nonlinear and complicated. Being able to take sensor data and understand what it means and how to use it for any type of control requires a much, much deeper level of technical understanding of what is going on in the chamber. Plasma processing is one of the core enablers for the entire semiconductor industry, but it has a level of complexity beyond anything else that happens in the fab. And that is where we are really focused, and why plasma simulation is so critical. This is also one of the benefits of the acquisition of Esgee Technologies and the capabilities it adds in chamber-scale plasma simulation.

SE: So there’s a lot of data to crunch?

Fried: Yes. We generate substantial amounts of data across the value chain. System-wide, being able to generate data that’s actually useful and actionable is critical. This data speaks perfectly to the core differentiation in our equipment. We are not generating data for the sake of generating data. We’re generating data that’s going to be immediately actionable in our core equipment business.

SE: What’s the impact on yield, and time to yield, if you can predict all these things ahead of time?

Fried: If we can understand these effects as we develop the equipment and process recipes, we can deliver better capability at higher quality and at lower cost. So if we can deliver equipment and processes to a fab at better uniformity, lower variability, and with improved capabilities, then the fab can incorporate those processes into their flows in a more robust way and design for higher yield. Every time we scale a node, the defect size and density must go down by significant factors. ‘Good enough’ gets much more difficult at every scaled node. Every chip has more transistors, more logic, and more complexity on it. Designing equipment and unit processes that have higher quality and better capability is critical in scaling to the next semiconductor node. Fabs can work with some level of defectivity, but with the next generation of chips they will have to support lower defectivity than they can now.

SE: Will machine learning help to interpret this data?

Fried: Absolutely, and interpreting this data will require highly complex simulations that will include machine learning. We’re going to be generating a large amount of simulation data in addition to the hardware data we generate in our labs. We can combine all of that data using advanced data analytics, machine learning, and AI to drive even faster simulation and a faster time to solution. This is part of what we are working on, along with a number of other initiatives.

SE: What else do you see as benefits from simulation?

Fried: Using simulation, combined with physical infrastructure, we can deliver solutions to our customers and the industry faster and more cost effectively than we’ve been able to do without simulation. This is a new tool, and it’s a really big hammer that also connects with other digital capabilities. As an example, for someone who doesn’t have a fab — such as a university — they can simulate innovative ideas just like you might do using actual semiconductor processing hardware. This lets us open innovation channels worldwide. If you think about semiconductor education in the developing world, how do you achieve that? They don’t have fabs, labs, or semiconductor equipment to enable exploration. But if we can provide the digital equivalent of this equipment, we can start tapping innovation from people around the planet who have never had access to this equipment in the past. Previously, this innovation originated in a lab or a fab, but there are not that many of them available worldwide. We can digitally transform this industry to allow people from all around the world with different socioeconomic backgrounds to participate in this innovation. This could lead to an explosion of innovation in our industry.