AI-based Scheduling Software to Upgrade Chip Fabs
Flexciton points out that chip fabs are the most complex manufacturing environments in the world, yet they’re slow to adopt artificial intelligence (AI) software that has the power to make them smarter by automating decisions and optimising production.
Flexciton CEO and Co-Founder, Jamie Potter, explained, “Fabs commonly use heuristic software where rules have been written by engineers to run the production and are based on historical data. This set of rules is used by the software to dispatch the wafers to the various tool sets across the fab. There is no intelligence built into the software. The intelligence comes from the skilled industrial engineers who use their years of experience to adjust and tweak the scheduling rules to allow for changing conditions and production requirements”.
A truly smart fab uses scheduling software with intelligence built-in that can holistically look at the current state of the production to spot bottlenecks and idle machines to automatically adjust to prevent them. Flexciton software already knows how to operate a fab, so it does not need the time-consuming, rule-writing phase for a new fab installation or the replacement of existing scheduling software. It has the intelligence and pre-programmed knowledge to look at the data from all the tools and work out how to run them effectively and efficiently, all in real-time.
The data that both approaches require is the same and is already automatically gathered by most fabs and stored in their MES. The heuristic approach looks at the current data using on-premise computers and generates the schedules based on the rules. “We have found that this is not very efficient as it is based on historical data and pre-set rules,” he added. “The rules have to be written by engineers and so they are dependent on people’s knowledge and experience to try and anticipate what will be the right decision in all possible scenarios. Because it is impossible to predict all instances that may happen in the future, the rule set requires constant tweaking and maintenance. This is a huge amount of work and requires constant updating with new rules and, if the engineers that wrote them leave – you’re in trouble. This is where heuristics are particularly weak. The rules aren’t necessarily wrong, but the decisions they give are inconsistent. They also struggle to cope with most complex constraints and are obviously unable to automatically adjust to rapid changes happening in a fab. These decisions aren’t usually overridden, they are just allowed to happen in the fab at the expense of factory efficiency that can steadily decrease over time.”
For example, the rules don’t take into account a hard-physical constraint in the factory and, as a result, may try to provide an instruction which is physically impossible to implement. Then an Operator must step in to override the system.
