The system conductor

Eindhoven University of Technology, ASML and ESI (TNO) are working on LSAT, a model-based tool that bridges the gap between systems engineering and the mono disciplines. Especially during the first design phase, this is of the utmost importance for the optimization of the overall system behavior and the supervisory controller. Through formal methods and domain-specific languages, LSAT quickly provides insight into which architecture will be most productive. The three partners hope to set a new industry standard with the tool.

Flexible manufacturing systems can be seen everywhere, from the semiconductor industry to car factories. What these systems have in common is their complexity. This is often because they have to be able to adapt to different part types, strict timing requirements, a large number of production steps and a platform with many shared resources. For the process to run smoothly, all those modules and components have to play together like a symphony orchestra.

The performance of such a flexible manufacturing system is highly dependent on the configuration and conducting capabilities of the controller. You won’t achieve the highest possible output just by putting your faith in the knowledge and skills of the system architect. It simply takes too much time to manually calculate all design options. Moreover, it’s extremely difficult to make solid statements about unique systems that have never been designed before, purely based on past experiences and a series of set rules.

About ten years ago, Eindhoven University of Technology (TUE) and ASML put their heads together to tackle this problem. Later on, ESI and ICT-NL worked on the implementation of a tool and the research was expanded in the ESI Concerto and NWO RCPS programs, and at the European level in Arrowhead Tools. The result is a model-based development tool with which designers can specify system behavior via domain-specific languages and optimize them at an early stage for the throughput of the final machine. The package is called LSAT, which stands for Logistics Specification and Analysis Tool.

“LSAT is a tool that allows you to quickly explore all design options and discover in the first phase of the development process which system configuration will yield the highest throughput,” says Jeroen Voeten, professor of cyber-physical systems at TUE. Wasn’t there already a solution in the large toolbox available to designers today? “In system control, you have different layers. Mechatronics typically deals with movements from A to B. Above that is a layer for coordination and above that a layer for planning. The higher you go, the less standardization there is. Most companies do have a supervisory control layer, but they all have chosen a different approach. And that means there are no uniform modeling tools that can deal with this problem.”

“One of the best-known tools for model-based design of dynamic systems is Simulink from Mathworks,” Voeten continues. “Very widespread in the mechatronics world because it allows you to optimize for things like speed and accuracy of individual robot movements. LSAT hits a higher level where it’s about the coordination of all these movements. That’s really something different. It’s about productivity and all the design decisions you make to optimize it.”

This optimization is done with formal methods, Voeten explains. “These are mathematical techniques with which you can provably determine the best solution. That’s a step further than simulation, in which you only play out one system configuration. With LSAT, you look at all possibilities and calculate which one will get you the best result.”

Optimization with formal methods is a subject perfectly fitted for the researchers at TUE. After all, it’s a fundamental technology for which you only need to have limited knowledge about the domain for which you’re developing it. Within the multidisciplinary partnership, this topic is entrusted to the scientists in Eindhoven.

ESI is closer to the industry and focuses on the domain-specific languages (DSLs) needed to describe the design in jargon that LSAT understands. “In a generic language like Java or Python, and also in Simulink, you can write anything you want – as a designer, you have enormous freedom,” says Bram van der Sanden, a research fellow at ESI. “But the more possibilities you have, the harder it is to properly analyze the result. Within LSAT, we work with a domain-specific language, in this case focused on flexible manufacturing systems. To give an example, in ASML’s wafer handler, engineers talk about wafers, loading and unloading robots, and setpoints. These are typical model elements that you don’t see in a generic language but are normal in a DSL. As a result, a designer can use the jargon from his day-to-day work and doesn’t have to become a software expert. Even non-specialists can quickly write down what they want to achieve because it’s in the same language they speak every day.”

A DSL contains a number of architecture rules. “You can describe exactly what is and what isn’t allowed,” says Van der Sanden. “When drawing up the model, users receive much more support because they get immediate feedback if they do something that’s not allowed. And because they can set up their model much faster, they can cover a much larger design space in an efficient way.”

To provide insight into the application of LSAT, Voeten and Van der Sanden, together with co-researchers Yuri Blankenstein from ESI and Ramon Schiffelers from ASML, use the fictitious Twilight system, which represents a very simplified version of a wafer handler. Twilight consists of four main parts: a loading robot and an unloading robot that transport balls, a conditioning station that heats the balls to the correct temperature, and a processing station that drills a hole in a ball. The balls must pass both stations in the correct order and of course, the robots must not collide.

“To arrive at the optimal configuration and the highest throughput, we need to do two things: model and analyze,” explains Van der Sanden. “First, we model the platform: we specify the peripherals such as grippers and motors. We then assign these to system resources.”

All movements are then given a profile so that LSAT knows exactly how much time it has to allocate for them. “These are the building blocks with which we can model all kinds of actions in the system”, says Van der Sanden. “For example, to get a ball from point A to point B requires a whole recipe of partial actions that have to be performed successively or simultaneously.”

When all actions have been entered, LSAT determines the optimal order for the best possible product flow. This is how the system will have to be programmed. The tool also displays the result visually in a Gantt chart. Van der Sanden: “This shows you the critical path and the resources that might form a bottleneck. Within Twilight, that turns out to be the loading robot. If you want to speed up the process further, it might be wise to take a look at that part first. Maybe you can choose a faster robot. Or you may need to introduce a third robot into the system. The great thing about LSAT is that you can quickly calculate such an alternative by updating the model. The building blocks are already in place.”

Twilight is a relatively simple and small system, but LSAT works just as well for very complex machines. For example, ASML has used the tool in the development of the wafer handler that processes the flow of silicon wafers to and from its lithography machines. With LSAT, the developers in Veldhoven evaluated what effect changes to the mechanical platform and the supervisory controller had on the system performance.

Voeten: “You can also use the LSAT model as a specification. Those are often written in Word files, but you can’t execute them and they’re full of ambiguities. The LSAT model is an excellent blueprint for the rest of the development.” This is precisely what ASML has done in its collaboration with VDL ETG. To develop and build the wafer handler, VDL ETG used ASML’s LSAT models to create UML diagrams and generate code.

At ITEC, developers have applied LSAT to model their die bonders. This gave them insight into the critical path, budget and productivity. “Modeling systems in LSAT leads to more complete specifications because the dependencies between components must be made explicit,” said Sam Lousberg, mechatronics engineer at ITEC.

“LSAT does scale up to very large systems, but there’s a limit to the state space that the tool can handle,” Voeten admits. “In general, you can easily solve this by constraining the state space manually – for example by defining more boundary conditions – or by dividing the process and optimizing each piece separately.”

Voeten believes that LSAT is an ideal multidisciplinary tool. “All machine builders who want to improve the logistics in their machine and increase throughput can get started with LSAT,” he says. “It’s a lightweight tool that doesn’t require every company to reinvent the wheel. We’ve already seen that it works in the ASML domain and the VDL ETG domain. The assessment at ITEC gives us a lot of confidence that it’s more widely applicable.”

ESI is currently working hard to make the tool available in open source through the Eclipse Foundation. Van der Sanden: “We want to move towards an ecosystem where companies can make their own additions, and where we as researchers can launch new optimization and analysis techniques. A joint tool that everyone can benefit from.”

The ambition is even greater. Voeten: “We’d like to see LSAT become the industry standard. We have all kinds of extensions on our roadmap for this. Ultimately, we want you to be able to specify the entire supervisory control layer, including exceptional behavior, and then generate code directly and automatically from the LSAT models. That’s the dream. We’ll need another ten years before we’ve achieved that though.”