Postdoctoral researcher – Energy‑efficient AI diffusion models

About the role

The project investigates how emerging analog and photonic hardware can accelerate diffusion‑model inference while keeping energy consumption low. You will bridge mathematical convergence theory with device‑level non‑idealities to quantify and optimise the trade‑off between energy use and sample quality.

Key responsibilities

• Make the score error measurable on Gaussian targets and report the induced L2 score error.
• Map device physics (PCM, FeFET) onto the theoretical error budget, separating proven from measured components.
• Extend existing convergence guarantees to mixed error models (part frozen, part stochastic).
• Co‑design with INL to deliver a proof‑of‑concept energy‑accuracy frontier for a PCM‑based photonic matrix‑vector engine.

Required profile

• PhD (or defended) in applied mathematics, statistics, machine learning, electronic engineering, computer science or a closely related field.
• Strong foundation in probability, stochastic analysis, numerical analysis, optimisation, or machine learning (e.g., SDEs, Girsanov’s theorem, diffusion models).
• Familiarity with convergence theory of sampling algorithms or ability to acquire it quickly.
• Fluent English; French not required.

Required skills

• Python programming, with experience in PyTorch or JAX.
• Machine‑learning modelling, especially diffusion or score‑based generative models.
• Numerical analysis and optimisation techniques.
• Probability and stochastic analysis.
• Device modelling and circuit simulation (photonic, ferroelectric, analog in‑memory computing).

What we offer

• Double supervision with weekly meetings between ICJ and INL labs (both on the same campus).
• Opportunity to publish in both mathematics and hardware design venues.
• Involvement in INL’s co‑optimisation projects and access to state‑of‑the‑art device expertise.
• Support for conference travel.