A general formula for deep learning success in semiconductor manufacturing

Abstract

This is a management overview of our experience in how to apply deep learning for semiconductor manufacturing projects. Deep learning is a transformative software technique, largely enabled by "useful waste" that is now possible with Peta-FLOPS level computing available with GPUs. DL has achieved many "firsts" that tens of years of effort by the best computer scientists could not achieve before. But production deployment of DL projects in semiconductor manufacturing including mask manufacturing has been difficult to attain. There has been a number of successes reported at this conference and elsewhere, but a common theme in deep learning papers in our field is the lack of availability of a large amount of data. Deep learning needs lots of data to train with because it is a pattern matching technique. It needs to see enough patterns to recognize all the situations a production mask might contain. In addition, deep learning programmers improve the network by adding data to the training data set to disambiguate where the network is confused. it is essential to be able to add any kind of data at will quickly to improve the success rate of a deep learning network. In semiconductor manufacturing, real data is hard to come by because of confidentiality requirements and also because of the expense of generating masks and wafers. A key ingredient to the general formula for deep learning success in semiconductor manufacturing is to use digital twins to generate data at will. It is time consuming, resource intensive, and expensive to set up. But it is necessary to create a deep learning capability that can be deployed in production. We will conclude with an overview of the other necessary conditions for deep learning success.