As semiconductor manufacturing processes grow more complex and sophisticated, production defects become both more common and harder to predict. Traditional process control techniques such as statistical process control (SPC) are now too limited to reliably anticipate defects. In addition, production data is often fragmented and unbalanced due to the instability inherent to manufacturing processes. Different products, machines or even drifts on the same machine produce heterogeneous and inconsistent data.
Consequently, and despite the need for more advanced anticipation solutions, the penetration of artificial intelligence (AI) for quality prediction remains limited. This is especially true for more advanced AI techniques, such as neural networks, which perform better in modelling complex systems but require vast amounts of balanced data.
Manufacturers are therefore stuck with legacy solutions and, despite the tremendous progress being made in modelling techniques, have limited perspective over the implementation of a fully predictive management of their operations at a time when their profitability is increasingly impacted by this lack of anticipation.
Lynceus provides an innovative quality prediction solution that leverages cutting-edge AI techniques to deliver high predictive performance on sparse production data. SEMI spoke with David Meyer, CEO and co-founder of Lynceus, about how advanced AI techniques can empower semiconductor manufacturing.
SEMI: Can you tell us a bit about your role at Lynceus? When did your interest in AI and semiconductor manufacturing start?
Meyer: My co-founder, Guglielmo, introduced me to semiconductor manufacturing. For a researcher in AI, semiconductor manufacturing represents an ideal playground because of highly complex processes, tons of data and a strong appetite for predictive solutions, around yield optimization. The more we looked into it, the more we realized how relevant Guglielmo’s transfer learning expertise was for this industry, especially as we needed to leverage multiple sources of fragmented data and needed to comply with very high standards of performance.
While Guglielmo focuses on developing and optimizing our models, my role is mostly concentrated on planning and structuring our development as a company and setting things in motion, from commercialization to recruitment and financing to product.
We know how to keep ourselves busy!
SEMI: What industry challenges is your software aiming to solve?
Meyer: In semiconductor manufacturing, the scale and complexity of operations are such that defects which negatively impact yields are both common and very hard to detect before the final test. It is not uncommon to see 15% to 20% of wafers scrapped at the wafer fab level.
Lynceus software solves the problem by predicting the result of quality tests for each unit processed in real time. Our technology relies on both deep and transfer learning, allowing us to reach high predictive performance on sparse production data, and to maintain this performance over time despite process changes.
Concretely, we deploy the power of neural networks on problems previously inaccessible to machine learning, reaching 10 times the reliability of existing solutions. When we kicked off four proofs of concept (PoCs) with semiconductor manufacturers, the software reached unseen accuracy on production data and enabled us to convert our first customer.
SEMI: How would you describe your approach?
Meyer: Lynceus changed the status quo by providing a versatile quality prediction software that is reliable because it is designed to provide optimal and stable predictive performance in production environments, such as modelling complex systems such as machines. In cases where data is limited, the software leverages similar data sources to maintain performance despite changes in the production process. In addition, our solution delivers a simple signal for each unit in real time, indicating to engineers if the unit is defective.
A key advantage of our solution lies in the modularity it enables, which empowers manufacturing operations rather than create disruptions, hence focusing on ROI and actionability. This means that our signal can be used in several ways with varying integration requirements. The first application is, for example, data-drive sampling, where manufacturers use our signal to identify the units to test instead of sampling at random and thereby capture more defects. Beyond data-driven sampling, our solution can be used to monitor machine performance in real time or even replace physical testing steps with a virtual test.
Ultimately, every step of the manufacturing process can be controlled at scale from day-one deployment to enable both significant yield improvements and cost savings.
SEMI: How are deep and transfer learning applied to ensure quality prediction in semiconductor manufacturing? Are you able to efficiently deploy neural networks on sparse production data?
Meyer: Our solution is powered by the combination of two of the world’s most advanced AI techniques: deep and transfer learning. The groundwork for our technology has been laid by our co-founder and CTO, Guglielmo, who leverages world-class expertise in both fields.
The combination of deep and transfer learning enables the deployment of neural networks (the most advanced modelling technique available) on limited and unbalanced data. Deep learning is a modelling technique inspired by the functioning of the human brain whose performance in modelling non-linear systems, such as machines, is unrivaled. The brain is notably able to extract features and generalize in order to establish complex relationships among multiple parameters. Transfer learning is a set of techniques for optimizing the training of a deep learning model that enables similar sources of data to model a target dataset. It means that if we want to predict defects on a given product processed by a given machine, we can use data relative to other products and other machines to complete this task.
Transfer learning is therefore key to model sparse production data, helping to lower data requirements to achieve high predictive performance on a given use case through the use of adjacent data sources. It also helps maintain stable predictive performance over time by supporting changes in manufacturing processes, such as machine mismatch, drift or introduction of new products.
SEMI: Many semiconductor companies are already seeing the results of deep and transfer learning applications. Can you give us an example of a success story?
Meyer: We are currently deploying our solution in major fabs in both Europe and the U.S. for use in predicting quality for plasma etching, Chemical Vapour Deposition (CVD) and Chemical Mechanical Polishing (CMP) processes. Our first engagement with an Italian foundry is an interesting case study, as it was the very first customer we interacted with. It asked us to model a plasma etching process to accurately predict the critical dimensions (CDs) on an initial dataset of c.1k wafers. We managed to predict the CDs with a c.1nm accuracy, so they gave us more data relative to several other etching machines. We maintained this performance despite the high variability in input data (several machines, tens of different product types). We have since deployed our solution and it is being used daily by their engineers as a tool to identify defects and monitor the performance of the etching process.
And who are your competitors?
Meyer: Lynceus currently is the only solution to integrate transfer learning, enabling the deployment of neural networks on production data at scale. Even in the most advanced wafer fabs, engineers train dozens of models and update them every few days to adapt to their production data. We are the only company able to propose a single model that can scale between machines and remain reliable in case of machine drift or introduction of new products.
We do not have competitors yet, but we are ready to accept the challenge!
SEMI: How did the pandemic impact semiconductor manufacturing and what you do see for the future?
Meyer: The current situation has already strongly impacted manufacturing. From supply chain diversification to social distancing rules within factories, structural changes lie ahead, with both challenges and opportunities.
As manufacturers relocate some of their assets closer to their end markets, or implement tighter safety rules within factories, their operating costs will increase, which will prove challenging in a sector where margins are already low. Smart manufacturing solutions that enable further automation and improving yields can help them balance this impact by increasing efficiency and productivity. Every crisis is a catalyst for innovation, and I am convinced smarter manufacturing will see major leaps forward in the coming years!
See the Lynceus Product Demo.
David Meyer (right), CEO and co-founder of Lynceus, graduated from ESSEC Business School in France, and started his career in management consulting at L.E.K. before switching to operations. He led operational teams and defined, launched and scaled processes now supporting millions of rides at Uber and Circ.
Guglielmo Montone (left), CTO and co-founder of Lynceus, holds a master’s degree in Physics and a PhD in computer science from Università Federico II in Naples, Italy, then focused on the study of Transfer Learning in his PostDoc at Paris Descartes. There, he notably invented an architecture re-used by Google Deepmind. He has 10 years’ research experience at leading institutions and has published 12 papers.
The SEMI SMART Manufacturing Initiative is a global effort to promote awareness and interest about Smart Manufacturing with focus on delivering industry-recognized best-in-class programs and services to enable members to maximize product quality, productivity and cost improvements through Smart Manufacturing. Activities are focused on building out core capabilities to enable Smart Manufacturing across the microelectronics supply chain.
Serena Brischetto is a senior manager of marketing and communications at SEMI Europe.