Center For Deep Learning In Electronics Manufacturing: Bringing Deep Learning To Production For Photomask Manufacturing

The Center for Deep Learning in Electronics Manufacturing (CDLe) was formed as an alliance between D2S, Mycronic and NuFlare Technology in autumn 2018. Assignees from each alliance partner worked with deep learning (DL) experts under the leadership of Ajay Baranwal, director of CDLe. The CDLe’s mission was to 1) turn DL into a core competency inside each of the companies and 2) do DL projects that make some differentiating feature for products or services for each of the companies.

Five years later, CDLe has completed its mission by graduating more than 20 students from the alliance members and completing 30 DL projects, many of which are headed to or already in production. This contrasts with the opinion of industry luminaries in the 2023 annual eBeam Initiative survey in which only 22% of respondents thought that DL would be used in production in any stage of production mask making by 2023. CDLe and its participants have had more success with DL than the industry at large because of their approach: using digital twins to bridge the data gap that keeps DL prototypes from moving to production use and following a “recipe for DL success” that includes executive commitment, dedicated computing resources, and DL and domain experts.

Bridging the data gap with digital twins

Aki Fujimura, CEO of D2S, provided executive sponsorship for D2S at CDLe. In several papers, presentations and videos, Aki explained that a prototype is easy, but production is hard. Aki interviewed a 10-year-old DL student about a DL prototype he had constructed quickly for identifying cats. The interview made the point that the prototype is easy, but production is hard, particularly because you need lots of the right data for the results to be useful in a production setting.

Data is absolutely critical for training DL networks, but data is hard to get in the photomask industry. Mask makers have a lot of data, but they can’t give it to vendors and often there’s not enough of the anomalous data necessary production-quality DL results. Digital twins – digital replicas of actual physical processes, systems, or devices that can be used in simulations – are used in simulations to create the right amount of the right kind of simulation data to train DL networks successfully. This is the only way to reach the stringent quality requirements of the photomask industry.

The recipe for DL success

Ajay and Aki have been thought leaders in the mask industry for using DL and have shared their experiences freely at various conferences. Many of these talks or papers are available at cdle.ai. The recipe for success as summarized by Ajay and Aki in one of these talks includes these ingredients:

• Executive sponsor with a long-term commitment. Along with Aki, Mikael Wahlsten at Mycronic and Noriaki Nakayamada at NuFlare provided executive sponsorship for their companies. Prototypes are easy but it takes a long time to get them into production. You need executive commitment to see it through.
• A lot of computing resources. You’ll need a million-dollar computing budget in order to build a production system with up to 95% accuracy. You also need it to use digital twins to generate the volume of data needed to train production-quality DL networks.
• Domain and DL experts. The hardest part is people. According to Aki, “Our CDLe experience was to turn a domain expert into one with DL expertise through immersive learning for 3-6 months. One of the best parts of how we set up CDLe, in retrospect, is that we coupled DL experts with domain experts from each company.”
• Immersion. “Just like new language acquisition, immersion where you forget other things and focus only on DL was a key to success,” according to Aki. The assignees at CDLe worked together in the same office, sharing DL experiences with each other and only worked on DL projects all day every day. What’s also clear through the history of CDLe is that you need to select the right project, a project suited to DL. This is one reason that DL is so often used in image recognition: it is very well-suited to that type of task. DL is a statistical method, so something where a prediction or an estimation is the objective is suitable. Something where human judgement might be 90% accurate is good for DL to automate the task with the same or greater accuracy.

DL from CDLe in production

Each of the CDLe member companies has developed DL applications that are now in production use. NuFlare and CDLe used digital twins built by CDLe to generate 850,000 data points to train a 12-million-node network to diagnose semiconductor variable shape eBeam (VSB) mask writer failures. The NuFlare DL application delivers 90%+ accuracy for VSB mask writer failure diagnosis. A SEM image of a result of a failure is analyzed by the DL application to provide a quick diagnosis.  For a fully utilized mask writer, it’s critical to solve the problem as fast as possible. The application helps engineers diagnose the problem and make a fix faster than before.

Mycronic chose to send a long-term assignee to CDLe combined with short-term assignees for 3 months to work on actual projects. One such project was to create autoencoders for detecting abnormal behavior in their flat panel display mask writers. For customers, it guides them to where inspection should be done. For service engineers, it guides them to where to concentrate their repair efforts. The CDLe projects helped Mycronic into the mind set of big data and what data needs to be collected in the future for new solutions for customers.

The final word goes to Ajay Baranwal, who was the director of CDLe for the entire five years: “All of us at CDLe were teachers, learners, researchers, and implementers. We maintained that same philosophy throughout. My sincere gratitude goes to each of the three partners and the executive sponsors for putting the structure, people and trust in place to create CDLe, then extending its mission to five years.”