Job description
Join the NWO Perspectief FIND program and develop methods to adapt Transformer-based foundation models for defect detection where data is scarce and unlabeled. Explore few-shot learning, self-supervised adaptation, and synthetic data generation to enable robust, scalable AI in semiconductor and printing systems. Work with leading industry partners like Canon and help transform quality inspection in next-generation high-tech equipment!
Information
Industrial edge deployments—in semiconductor manufacturing, industrial printing systems, automotive radar, smart mobility cameras, and HealthTech—require on-device AI to ensure low latency, privacy, and resilience. Today’s Transformers models scale poorly and assume abundant cloud resources. The research program FIND aims to deliver architectural and algorithmic breakthroughs that enable foundation models to run predictably and efficiently on embedded processors and accelerators.
FIND is a research program funded by the Dutch government and industry that brings together 5 universities, 11 companies (startups to multinationals), and 2 knowledge institutes to develop foundation models (large AI models) for Dutch high‑tech industry, with strong emphasis on edge deployment, privacy, and timely decision‑making. Partners include ASML, NXP, Canon Production Printing, ASMPT, Technolution, Signify, Shell, Stryker, TNO, and others. A total of 12 PhDs will be employed on the FIND program covering topics from foundation model pre-training and multimodal adaptation to architectures and compression for edge deployment while targeting real-world validation in domains like HealthTech, smart industry, and autonomous mobility.
This PhD position focuses on adapting and fine-tuning Transformer-based foundation models for defect detection in high-tech manufacturing environments where only limited and largely unlabeled defect data is available. Current solutions typically rely on supervised CNN-based models trained on large labeled datasets, which fail when defects are rare, vary across machines, or when labeling is prohibitively expensive. These approaches lack flexibility and generalization, making them unsuitable for dynamic industrial settings with scarce and imbalanced data.
You will also explore few-shot learning, self-supervised adaptation, and multimodal integration techniques to overcome data scarcity and improve robustness. Unlike existing methods that depend on exhaustive annotation or handcrafted features, this research will leverage the rich representations of foundation models and develop strategies for zero-shot or few-shot adaptation. You will investigate domain adaptation, synthetic data generation, and cross-modal learning to enable models that generalize across defect types and machine configurations. This ensures scalable, accurate defect detection even in low-resource industrial contexts.
The resulting models will be validated in collaboration with a lead high-tech company, demonstrating how foundation models can transform quality inspection by reducing dependency on labeled data and enabling rapid adaptation to new defect patterns—closing the gap between AI capability and real-world manufacturing constraints.
Research group and company
This position is embedded in the Mobile Perception Systems (MPS) Lab and Electronic Systems (ES) group within the Electrical Engineering department at Eindhoven University of Technology (TU/e). The MPS lab and ES group have a strong history of collaborative research projects leading to real-world impact.
This PhD project is executed in close collaboration with Canon Production Printing which is a global leader in high-end digital printing, offering advanced hardware, software, and services aimed at professional and industrial-scale print environments.