Back-End Automation Tackles Growing Complexity

Experts at the table: Semiconductor Engineering sat down to discuss back-end automation challenges in advanced packaging with Michael Lowman, senior product marketing manager for Data Analytics at Cohu; Aftkhar Aslam, CEO at yieldWerx, Woo Young Han, product marketing director at Onto Innovation; and Lihong Cao, senior director of engineering and technical marketing for ASE. What follows are excerpts of that discussion.

L-R: Cohu’s Lowman, YieldWerx’s Aftkhar, Onto Innovation’s Han, ASE’s Cao.

SE: As advanced packaging moves toward chiplets, 3D stacks, and fine-pitch interconnects, how is back-end automation evolving?

Lowman: We make test equipment — basically test handler equipment and different aspects of the test process and finishing — and the majority of our challenges are related to thermal management. Today’s devices require better thermal performance during test, and they’re tested at high power — up to 3,000 watts. Those devices are expensive and precious, so customers need in-line inspection for the handling equipment before and after test. That allows us to ensure the device is not damaged, and we can react in real-time to avoid a cascading problem. I believe there are other challenges coming that haven’t hit us yet, like equipment cleanliness and stable device control during test. Those will be interesting areas and challenges moving forward.

SE: Do you feel like the wide variety of packaging is creating challenges for automating the whole thing?

Lowman: Yes. The device size alone creates challenges. They’re getting bigger and taller, and there are all sorts of challenges associated with that. What used to be picked up with a single head now might require two heads, and there are mechanical constraints associated with that.

Aftkhar: We’re on the tail end of this, and we collect all the data. There’s complexity in picking and actually testing the parts. Where we see the challenges is that those parts have to be inspected. Fortunately, there is new equipment coming out in the market, like optical inspection tools, X-ray tools, and acoustic tools that look for warpage and other kinds of defects, together with lab tools like FIB-SEM tools. The real challenge we see is that each of these tools is producing data in different formats and using different coordinate systems. If you’re looking at chiplet XYZ, how is that translated to the actual major package itself if the data is in different formats? Customers are saying, ‘I have multiple islands of data that come from these tools. How am I supposed to join them?’ They’re looking for a single backbone where they can pull this data in and stitch it together. We also need to sequence it from a timeframe perspective, because you have different parts going into these complex packages. Where did the actual problem start? What was the root cause? Then trace the root cause back. Was it an assembly and test issue? Was it pure testability? Was it something done in the fab all the way back to design? Then turn those into action, saying, ‘I found the root cause. I need to fix it in design, in manufacturing, in my test programs, or my actual manufacturing assembly line.’ Those are the challenges we see. Customers are tired of seeing islands of data with no uniformity across those islands.

Han: We make inspection and metrology equipment, and the form factors are changing. In the past, we used to inspect a single wafer, but now we’re seeing up to triple-bonded wafers. With three wafers bonded, it gets over two millimeters thick. With all these wafers bonded, there’s a lot of stress added, and the new warpage requirement we’re hearing is up to plus-or-minus three millimeters. When you have a two-millimeter-thick substrate with a plus-or-minus three-millimeter warpage, it’s a whole different ballgame than traditional wafer handling. We’re hearing about RDLs getting below a micron, and with all these high-value bondings and fine pitch, trying to do sub-micron inspection with these kinds of warped surfaces with high contour variation has a lot of challenges. Another trend we see with chiplet packaging and HBM is that it needs wafer-level molding. With wafer-level molding, that’s another layer of challenges for inspection and metrology. There are new types of defects created during molding, including edge and backside defects. This creates a lot of yield-killer issues with advanced packaging. Trying to accommodate for this warpage and for thick wafers is adding all new requirements for us.

Cao: From ASE’s point of view, even though we say ‘back-end,’ we really do middle-line offline as well — for chiplets, RDL, substrate, and board-level assemblies. I’d like to talk about this from two big areas: handling and inspection. From a handling point of view, there are two dimensions to consider. First is X-Y. The module size is becoming larger and larger, whether it’s a single wafer or multiple chiplet wafers when we do molding. We already have two-reticle size, four-reticle size, five-reticle size. How do we handle that type of wafer-level size? On the substrate side, people are already talking about 100 x 100 millimeters to 150 millimeters. Some are discussing 170 millimeters, which is already beyond current capabilities. How do we handle x, y dimensions during both process and test? The second dimension is z — warpage. We’re seeing larger warpage right now — more than machines normally handle. It used to be maybe one millimeter of warpage. Right now, it’s increasing to three millimeters, and we’re even seeing more than three millimeters in Z-height. Warpage combined with a large body size is already a big challenge.

We’ve also found thin wafer handling to be a very big challenge. When we do the middle process and thin down the wafer, right now people are working with 50 microns and 100 microns. All this thin wafer handling during the process is a big challenge, and we definitely need automation because of particle and contamination requirements. With very fine line spacing — two microns by two microns and even less — particle control requires automation. That’s just during the process. For metrology and inspection, we need high-resolution inspection tools — in-line tools and IFA tools. Those need to be integrated with the handling equipment, with the process machines, and with the automation systems. How do we achieve that integration? That’s the challenge, along with maintaining throughput. We’re investigating AI to help consolidate all these requirements.

SE: What does a practical closed-loop system between fab, assembly, and test look like today? What’s still missing to make automated, real-time decision-making more the norm than it is now?

Aftkhar: In Yieldwerx’s definition of closed loop, we go all the way from design to manufacturing, from manufacturing to wafer sort and wafer sort testing, then to final assembly and test all the way down to packaging, PCB board testing, and final end application. Not many companies can have the data for that loop. If you’re a fabless company, you don’t have access to fab data. You only have access to design data, and assembly and test data. If you’re an OSAT, you don’t have access to manufacturing or design data sometimes. The perfect goal is to assume we have access to all that data, from design to manufacturing to assembly and test to the edge application. We’re looking at feed-forward and feed-back. Let’s start with feed-back. I get to assembly and test, or wafer sort testing, I find a problem. The first thing I’m going to look at is whether it is a test problem. Is it to do with my hardware, my instruments, my probe cards, or load boards, as an example? Then I start to look outward from there and ask, ‘Does the wafer data that the foundry or fab gave me point to a problem? Then I go up the food chain to the fab — deposition, etch, and furnace data. But it goes beyond typical metrology data to say, ‘Which lot was the wafer in? Which chamber was it? What was the recipe? What was the state of the machine?’

This becomes a multi-dimensional problem, and there is a ton of data. The lot numbers change. Lots get split. Lots get combined. One of the tools we have is what we call lot genealogy. We can trace data all the way up the stack from final assembly and test, to design, and vice versa. We’ve developed this correlation engine that, assisted with AI, can pretty reliably pinpoint where we believe the problem lies. Now you have to train the models to ask, ‘What does good look like versus what’s bad and what’s maverick?’ You don’t want an engineer spending that time defining what’s positive and what’s negative. We’ve trained the models to look for good data versus bad data. We compare material coming off the floor and say, ‘This is bad based on these signatures.’ Sometimes those signatures don’t point to one single root cause — they point to multiple root causes. The challenge is being able to join the data in a meaningful way to do this correlation and feed-back. But then the power becomes more powerful for feed-forward. In the manufacturing process, if I’m at step X, I can analyze the data and ask, ‘What is the impact going to be on subsequent process steps all the way down to test?’ Then we can control how those wafers and lots get tested by either loosening test limits or tightening test limits, or doing adaptive probe where I’m only probing a certain area of the wafer and screening the other material that’s going to be discarded anyway. These are the challenges and solutions customers are looking for in a total closed-loop solution. Very few companies have done it, but the challenges all go back to the same themes. What’s my data format? Am I able to join the data together with good business rules to then deliver closed-loop solutions?

Han: We’re getting a lot of new requirements from our customers for data integration and AI data analysis. Rather than trying to use a traditional Windows-based file transfer system, we’re hearing a lot of new requirements that are more machine learning-friendly to analyze data. We’re getting requests to use a new framework such as gRPC — a Google-based file transfer system — and a lot of new ways to send wafer inspection results and images so that for our customers it’s easier to analyze all these inspection methodologies and analyze the images through machine learning and AI-based systems. That’s a new requirement and requires us to develop software. As an industry, inspection and metrology equipment vendors are tailoring toward those needs. Our customers are working with AI companies to make their lives easier to analyze the results.

Cao: We have levels of challenge between upstream and downstream processes. We receive wafers or design files from different customers and foundries. We also receive different wafers from different foundries. Even though we already have a very good system for data login and communications, we still feel that because there’s no standardization, we receive data in different sizes or with different data requirements. We try to use our standard internally in ASE — we have our internal closed-loop system — but we still try to manage different sizes from outside. We feel slightly challenged there, and that’s still ongoing.

If you talk about closed loop inside ASE, we also have to communicate. We manage many different products and different advanced packaging technology requirements and different test-side requirements — in-line and also in our in-line analysis, and also for yield and different processes. We have different traceability systems set up. We applied AI to help us simplify the management of all these in-line data sites, as well as yield down to the test, because in our process we also separate the test in another site or another building. How do you make sure all the data translations, the data logging, and the mappings are installed and transferred correctly? We have software we developed along with traceability systems. We use machine learning for certain in-line data analysis to help with the test team and provide feedback to the labs for IFA (in-fab analysis). This forms the closed loop for yield inspection and yield learning and yield control.

Lowman: I like the idea of starting practically — what’s the practical first step — because a lot of the talk here is about what data is available and how do you format that data. Similar to what Wu Young was saying, we’re getting requests to provide data that’s going to be friendly to these analytics and machine learning applications. A reasonable first step is to pick a closed-loop system to start with.  Maybe test and finishing, for example. Then you keep the data collection within the same solution, maybe the same vendor. Then you start to learn through learning cycles of a closed-loop system there. You get rid of the data challenges or the diversity of data, and then you learn. Then you expand, and maybe you can generate some standards and start attacking the heterogenous data challenges. I haven’t been exposed to much closed-loop activity of this sort in the back end.  My thought is to start moving in some direction, keep it as simple as possible, and then learn.