The company was founded by David Meyer, CEO, and Guglielmo Montone, CTO, and is using advanced AI techniques to use early and sparse data from process monitors to improve yield and minimize chipmaking equipment downtime.
The company's business model is to provide a software product for fab-owners to operate and benefit from, Meyer said. So far the company has applied deep and transfer learning to etch, CVD and CMP processes, but the objective is to build an end-to-end yield predictor, he added.
Meyer said the approach appeals to chip manufacturers because Lynceus is able to model the relationship between process parameters and test results without requiring any changes in the production line or in testing protocols.
Guglielmo Montone, CTO, and David Meyer, CEO, of Lynceus. Source: Lynceus.
"Our initial results are positive – we predicted critical dimensions in a plasma etching process to within a 1nm accuracy – and we are now running a full set of experiments to benchmark the scalability, robustness and accuracy of our solution versus a broad range of existing modelling techniques, such as neural networks with fine tuning, neural networks alone, SVM [support vector machine], Random Forest," said Meyer.
Meyer's partner, Montone, has ten year's research behind him specializing in Transfer Learning and invented an AI architecture used by Google DeepMind for autonomous driving. Transfer Learning is an approach to storing knowledge gained while solving one problem and applying to a different but related problem and can be key to working with efficiently with sparse data.
This means that working with sparse inputs, which might come from a processing machine and from the broader environment, the software can indicate what process tweaks may be needed to keep the machine operating in specification. This can reduce both downtime and the time taken when a machine is down. And with the tremendous value inherent in the work-in-progress, direct improvements here are highly valued by fab owners.
Next: Heading towards 28nm