The software toolbox: Building scanner metrology software for a fleet of lithography hardware

Lithography systems are more than just hardware. Imagine a modern automobile without engine management software – all of the parts would be there, but the vehicle wouldn’t even start. Similarly, the machines that make microchips need software to carefully orchestrate the printing process.?

The software embedded in ASML’s lithography systems does more than execute a fixed set of chip-printing instructions. It takes real-time information about the printing process and uses that information to optimize the speed and quality of chip production. Our engineers are tasked with continually improving the software so it can efficiently handle even the most complicated microchip structures.? ?

Over many generations of lithography systems, however, repeated updates to our scanner metrology software turned it into an unnecessarily complicated tangle of algorithms. A few years ago, we decided to break the cycle. This ‘clean break’ also allowed us to introduce newer technology, such as model-driven engineering (MDE) tools. Now that we’ve rebuilt the software architecture with the capability to develop across platforms ‘baked in’, we’re in a better position to develop and deploy new features without impacting our customers’ bottom lines.??

Software’s role in scanner metrology?

The scanner system is the heart of a lithography machine. It’s where microchip patterns get printed on silicon wafers.??

Using measurements to deal with inaccuracies that affect patterning is known as scanner metrology. The software that enables scanner metrology takes in information about the system and the wafer and figures out what adjustments are needed to optimize the pattern quality.?

Currently one of the biggest challenges for software engineers working in my department is getting accurate and repeatable maps of wafer surfaces. These maps ensure that each new printed layer is in focus and properly aligned with the ones below it.??

Wafer mapping isn’t a new problem. But now our customers want to make deeper and deeper structures, particularly for 3D memory chips . The lower layers of those structures still have the same smooth 1–2 nm features as thinner chips, and on those nearly flat surfaces we are continuing to improve the speed and precision of wafer mapping by our cutting-edge extreme ultraviolet (EUV) and immersion deep ultraviolet (DUV) systems. At the higher layers, though, the vertical variations can reach tens of microns. Like a finely tuned race car trying to drive off road, our systems aren’t yet optimized for those conditions.

Consistency breeds complexity?

Our software must continually improve to handle the increasingly complex structures and creative designs of modern microchips, the demands for tighter tolerances, and the market demand for higher productivity. And when we update it, we don’t want the improvements to be available only on our newest machines. Lithography systems are built to last, and our older machines remain in heavy use. In 2021, for example, more than 90% of all PAS 5500 systems – the machines that kick started ASML’s success – were still in service. For TWINSCAN systems, the proportion is even higher.?

One of the biggest concerns for our software engineers is ensuring that updates don’t affect a lithography system’s behavior. A reliable chip manufacturing process drives the customer’s revenue, so any unexpected change – even an improvement – comes at a cost. To keep the customers happy, our software developers took no chances. In the past, whenever an update was needed, we kept and duplicated the algorithms that worked in previous software packages.?

While this way of building software was OK during the earlier product introductions, about seven years ago we realized that it was no longer tenable. For example, one of the functions in our code had gotten so long that its cyclometric complexity – a measure of how complicated an algorithm is – was over 300. (Ideally it should be less than 10, and our new code averages 1.2.) That meant the number of potential paths through the function was larger than the number of fundamental particles in the visible universe times the number of seconds since the big bang.?

Rebuilding software architecture?

Over the following five or so years we rebuilt our entire measurement sequence from the ground up. The first part of the refresh involved switching languages from quite an old version of C to a modern version of C++.??

The more innovative step was integrating MDE into our workflow. With this kind of software-development tool, the programmer describes the behavior they want the machine to display, and the tool generates the code to make it happen.?

Implementing code produced using MDE made testing our software a lot simpler. The tool formally verifies that running the code can’t cause any critical problems by, say, leaving the system in an inconsistent state or starting a measurement that never finishes. If the tool identifies any such problems, it doesn’t generate the code.?

We built our new software architecture with our entire fleet of lithography systems in mind. The old architecture had a lot of system-specific instructions. Now more than 95% of the measurement code is the same for all the machines. That consolidation makes it easier to deploy new features and updates.??

Still, if you ask a customer whose older system has been running reliably for years if they want to switch to a new system architecture, their initial answer will likely be no. It goes back to the question of consistency: Why would they risk changes in their machine’s behavior for a software update???

Many customers have bought system upgrades to boost their outputs during the recent chip shortage. The new architecture came with upgrades, such as layout independent leveling (LIL), a technique for faster wafer mapping that was developed for the NXT:1980Di scanner. This is now the default on all new EUV and immersion DUV scanners and on the highest-productivity dry DUV scanners. Many of the already-deployed lithography systems also switched over when LIL was released as an optional upgrade package and in maturity-extension packages.??

To convince the holdouts to switch to our new software architecture, we’re continually making improvements so it can deliver productivity and pattern quality that they can’t refuse.?

Now hiring: software engineers for scanner metrology positions

In our scanner metrology team, software engineers play an important role. They implement real-time computational models in C++ or Python, translating and fitting these into the wider software architecture of our lithography scanners. If you’re a software engineer interested in working in the fascinating field of scanner metrology, check out this scanner metrology software vacancy, or browse all software jobs. For those still studying, we also have a wide variety of internship positions.?