The posting
Interested applicants are invited to apply directly at the NUS Career Portal. Please note your application will only be processed if you apply via NUS Career Portal.
NUS Career Portal link: https://careers.nus.edu.sg/job/Research-Fellow-%28Physics%29/34838-en_GB/?st=56BB4E3AC27C24C4015CC62A82D5B3E3276B2B9D
We regret that only shortlisted candidates will be notified.
Job Description
About the Role The AI for Science Gym builds bottom-up AI capability across NUS science and engineering. Its Discovery Gym helps researchers who already know the basics of data science and ML use AI to push the frontiers of their own fields. Its Instrumentation Gym builds the shared data and model substrate those researchers work on. As Discovery Gym Lead for Spectroscopy you are an AI-enabled data scientist embedded in the spectroscopy community across the Faculty. You run the Gym for groups working in Raman and surface-enhanced Raman, mass spectrometry, NMR, infrared and related methods. Your two jobs are to translate their raw spectroscopic data into the Instrumentation Gym's shared substrate, and to help those groups apply semi-supervised machine learning to that data so they can build and own their own models. Spectroscopy produces abundant unlabelled spectra and scarce, expensive labels, so semi-supervised and self-supervised methods are the practical route to useful models, and teaching partners to use them is the core of the role.
Job description
- Weekly small-group peer-learning across the spectroscopy groups: participants work through hard problems on their own spectra, apprenticeship-style. Onboard each new group or department as it joins.
- One-month Discovery Sprints with spectroscopy partners: profile the raw instrument data before any modelling, catching detector saturation, axis drift, duplicate acquisitions and design confounding; agree goals, provenance and one shared preprocessing definition; deliver a working dashboard within week 1; then set research goals and go/no-go gates. Close each sprint by handing over the scripts and the reasoning so the group can run it without you.
- Translate spectroscopy into the Instrumentation Gym: convert each partner's raw spectra into curated, provenance-carrying datasets in the shared substrate, with documented preprocessing, acquisition metadata and known failure modes, and deposit them as tokenised pedagogical challenges with key annotations withheld. Contribute the spectroscopy-side workflow demos that let AI scientists work on this data without a spectroscopist present.
- Teach the semi-supervised toolkit: build baselines that work with few labels, including self-supervised pretraining on unlabelled spectra, pseudo-labelling and consistency regularisation, representation learning and tokenisation of spectra, and active learning that tells a partner which few samples are worth labelling next. Hold every reported number to permutation nulls, stability checks and a preprocessing sensitivity grid, and make batch and acquisition-day effects explicit before any claim.
- Run sprints that yield a pedagogical preprint, take part in scrimmages against competing AI meta-harnesses, and join the monthly Discovery Gym Lead methods-exchange and the weekly sync with the Architect and the other gym leads, including the Instrumentation Gym Lead whose substrate you feed.
- Working expectations: 1–2 active sprints at a time plus 1–2 in support; consultation ≤40% of your week; active projects wrapped within ~3 months; at least one paper or proceeding per year.
Qualifications
PhD in physics, chemistry, materials science, or a related discipline.
Experience & Skills:
- with a research record in spectroscopy. We want a domain spectroscopist who became a data scientist.
- Hands-on spectroscopy: you have acquired the data as well as analysed it, across one or more of Raman and SERS, mass spectrometry, NMR and infrared, and you know how a spectrometer produces a signal and where that signal goes wrong.
- Semi-supervised and unsupervised machine learning on real data: dimensionality reduction, clustering, classification with few labels, self-supervised or contrastive pretraining, pseudo-labelling or consistency methods, and an understanding of where each breaks. Fluent scientific Python with a modern ML stack (PyTorch or JAX, scikit-learn, pandas, numpy) and reproducible, version-controlled workflows.
- Experience with messy real-world instrument data, and the judgement to tell a promising dataset from a hopeless one, including the discipline to validate a result against permutation nulls and stability checks before relying on it.
- Ability to ship fast and visibly: a useful interactive dashboard on unfamiliar spectra within a week, handed over with scripts the partner group can run themselves.
- Excellent communication and teaching instincts; genuine fluency with AI-assisted development and the ability to teach it critically.
- A publication track record and the appetite to keep publishing.
Desirable
- Postdoctoral experience applying ML to measurement data, and electromagnetic or quantum-chemical simulation (for example COMSOL or DFT) in support of spectral interpretation.
- HPC or multi-GPU environments; containerised workflows; scientific data formats such as HDF5 or Zarr.
- Representation learning, foundation or self-supervised models, tokenisation of continuous signals, or physics-informed methods.
- Building or evaluating agentic/LLM workflows.
- Mentoring, teaching, or running workshops and competitions.
- Breadth across more than one spectroscopic method, or across spectroscopy and an applied domain such as biological, environmental or materials analysis.
