Postdoctoral Fellow - Generative AI Researcher - 3D Defect Synthesis

The Challenge Real defects in batteries, consumer electronics, and structural components are rare, expensive to induce, and nearly impossible to reproduce at scale. We are building the next generation of AI-powered industrial inspection — and the bottleneck is training data. The mission of this role is to develop generative AI models that synthesize geometrically and physically plausible defects directly into 3D CT and X-ray volumetric data — spanning the full physical scale from centimeter-level structural failures down to nanometre-level material anomalies — creating the synthetic datasets needed to train high-fidelity detection models without requiring real defective samples.

Detection scale targets: cm (10 ⁻ ² m) — Structural cracks, delamination, component failure mm (10 ⁻ ³ m) — Voids, inclusions, solder bridges, swelling µm (10 ⁻ ⁶ m) — Micro-cracks, porosity, dendrite growth 100 nm (10 ⁻ ⁷ m) — Interface delamination, thin-film defects nm (10 ⁻ ⁹ m) — Grain boundary, lattice-level anomalies What You Will Build Defect synthesis engine: Diffusion, GAN, or NeRF-based models that insert controllable synthetic defects into 3D CT volumes, parameterised by type, size, location, and severity.

Physics-aware rendering: Ensure generated defects respect X-ray attenuation physics, Hounsfield unit gradients, CT reconstruction artefacts, and material contrast — enabling real sim-to-real transfer. Multi-scale CT dataset pipeline: Tools to produce large-scale labelled synthetic training datasets from micro-CT, CBCT, and X-ray projections at multiple resolutions and across device types.

Closed-loop detection training: Connect synthetic generation directly to YOLO / segmentation / anomaly detection training loops, with real-scan validation benchmarks and iterative refinement.

Benefits

Internationally competitive, tax-free salary On-campus housing included Comprehensive health insurance Annual travel allowance Access to world-class CT, micro-CT, and imaging facilities on campus Vibrant international research community (100+ nationalities)

Core Requirements PhD in Computer Vision, Medical Imaging, Applied Machine Learning, or a closely related field. Hands-on experience with generative models — diffusion (DDPM/LDM), GANs, VAEs, or Neural Radiance Fields applied to 3D or volumetric data. Strong background in 3D CT or X-ray imaging: reconstruction pipelines, projection physics, or volumetric segmentation.

Experience building anomaly detection or defect detection models (e.g. Anomalib, YOLO, segmentation pipelines). Proficiency in Python and Py. Torch. Familiarity with MONAI, ASTRA toolbox, or SIRT/FDK reconstruction is a strong advantage. Advantageous Background Industrial NDT, materials science, semiconductor, or battery inspection experience.

Domain randomisation and sim-to-real transfer for detector training. Multi-scale imaging: micro-CT, SEM, FIB-SEM, or synchrotron data handling. HDF5 / Zarr data schemas and GPU-accelerated volumetric processing. Published work in generative models, synthetic data augmentation, or 3D reconstruction.

Application Questions (Required) All five questions below are mandatory. Please answer them in your own words. Generic or AI-generated responses will not advance in the process. Q1 — Debugging memory Describe a specific bug or failure in a CT reconstruction or generative model pipeline that took you more than a day to resolve.

What was the root cause, and what did you change? (We are looking for a real incident, not a hypothetical.) Q2 — Constrained format In exactly 3 bullet points — no more, no less — state what you believe are the three hardest unsolved problems in synthetic-to-real transfer for CT-based defect detection. Answers with more or fewer bullets will be disqualified.

Q3 — Literature opinion Name one paper published after January 2024 that changed how you think about 3D generative models or volumetric defect synthesis. Give the title, one thing it got right, and one thing you would push back on. Include the DOI or ar. Xiv ID. Q4 — End-to-end system experience Describe a generative AI system you built or contributed to that was used by an end user (not just a research prototype).

What was the input, what did the model produce, and how did you handle the gap between model output quality and what a real user actually needed? We are looking for evidence of practical deployment thinking, not just model training experience. Q5 — Code evidence Provide your GitHub or GitLab username and a link to one specific commit or pull request you are proud of.

In 2 sentences, explain the non-obvious decision you made there. Repositories must contain real commits predating this posting. How to Apply Submit all materials in a single application via the link below. Email applications will not be reviewed. CV / résumé (PDF, max 4 pages) Statement of purpose (1 page — why this role, why now) Answers to questions Q1–Q5 (in the form fields) Answers to Q4: describe your end-to-end system (in the form field) GitHub / GitLab link for Q5 Links to up to 3 relevant publications (optional) Applications are reviewed on a rolling basis.

Position open until filled.