Push multimodal GenAI beyond the lab. Join our research team as a PhD candidate to work on beyond the state-of-the-art model distillation and robustness methods, enabling efficient, reliable inference for challenging real-world problems in the semiconductor industry.
Industry context and motivation.Pretrained foundation models bring lots of potential in semiconductors metrology. Multimodal foundation models combining data from multiple metrology sources, enable high accuracy in the early stages of next generation manufacturing process, and across many different usecases. As the process matures and enters high volume manufacturing, these models can be specialized through efficient fine-tuning and distillation to improve throughput in terms of both compute and data acquisition, while maintaining guaranteed performance on critical failure modes.
This PhD project focuses on developing data-efficient methods for fine-tuning and distillation under strict (customer fab) data and privacy constraints while providing performance guarantees for specific downsteam tasks.
Customer data can be indeed highly constrained. In particular, customers might not allow data to leave the fab, and datasets that are available can be limited and imbalanced.
The model performance requirements are expected to evolve over time – from early research to high-volume manufacturing. ML models should be able to adapt to changing accuracy, speed, and defect-detection requirements across the full lifecycle of a process node, and should cover a wide variety of use cases, across different customers, with minimal adaptation.
Developing beyond state-of-the-art techniques enabling such ML behaviour of ML would have a profound impact – facilitating more robust metrology models and significantly faster time-to-recipe across the maturity stages of a semiconductor process node.
Research settings:
PhD candidate will be formally employed with the
Data and AI cluster at the Department of Mathematics and Computer Sciece and supervised by prof.dr. Mykola Pechenizkiy and dr. Ghada Sokar. The project is done in tight collaboration with dr. Jan Jitse Venselaar and dr. Jacek Kustra, of the ASML AI Research Team. The PhD candidate is expected to structurally spend time at both TU/e and ASML locations.
The project is part of NWO TTW Perspectief funded research program
Foundation for Industry (FIND) - Large AI models for a resilient high-tech industry providing further opportunities for collaboration.
The project has access to the national computing infrustracture, TU/e HPC cluster
SPIKE-1, ASML HPC cluster, ASML datasets, and potentially custom data through collaboration with e.g. IMEC.