CENTRALE LYON - Postdoctoral Researcher Position Deep Learning for Functional-Oxide Growth Video-to-Spectrum Prediction by RHEED / XRD Fusion

ecolecentraledelyon; Écully; Hybride; EUR 42 000 - 60 000; Plein temps

Résumé du poste

École Centrale de Lyon invites applications for a Postdoctoral Researcher to develop deep learning models that predict the final XRD spectrum of oxide thin films from RHEED growth data. The role sits at the AI–materials interface, with data from INL and LIRIS, aiming to deliver a real-time, calibrated spectrum predictor during MBE growth.

You will contribute to corpus structuring, probabilistic regression with uncertainty estimates, and a real-time demonstrator.

Qualifications

• PhD in AI/CS/ML or related field with strong publication record.
• Hands-on experience with video data and deep learning architectures.
• Proficiency in Python and ML frameworks (PyTorch or TensorFlow) are mandatory.
• Ability to work autonomously in a multidisciplinary team.
• Excellent scientific-writing and communication in English.

Responsabilités

• Develop and evaluate deep learning models to map RHEED video to final XRD spectra.
• Structure and curate versioned datasets combining video, growth parameters and ex situ spectra.
• Quantify predictive uncertainty and build probabilistic regression frameworks.
• Contribute to real-time demonstrator prototype during growth experiments.

Connaissances

Deep learning on video data

Python

PyTorch

TensorFlow

Bayesian / uncertainty quantification

Git and reproducible research

English communication

3D CNNs / temporal transformers

Probabilistic models

Formation

Ph.D. in AI/CS/ML or related field

Outils

Python

PyTorch

TensorFlow

Git

3D CNNs

Description du poste

PROJECT OVERVIEW

We are looking for a highly motivated Postdoctoral Researcher to develop innovative deeplearning models that predict the structure of functional-oxide thin films directly from their growth dynamics. Positioned at the interface between Artificial Intelligence and materials physics, the OXYD-IA project aims to design a deep model able to predict the final X-ray Diffraction (XRD) spectrum of an oxide thin film from the sole Reflection High-Energy Electron Diffraction (RHEED) video recorded during its growth by Molecular Beam Epitaxy (MBE).

The resulting tool will open the way to predictive, in situ control of oxide epitaxy - a process today dominated by a costly trial-and-error approach, in which the structural and functional properties of the films are only validated ex situ. You will join a genuinely multidisciplinary collaboration between the INL (Institut des Nanotechnologies de Lyon), which provides the operando experimental data and materials-physics expertise, and the LIRIS (équipe Imagine), which provides the deep video-learning and probabilistic-modelling methodology - both at École Centrale de Lyon.

A high-impact opportunity. AI for experimental physics is a fast-growing field, and OXYD-IA offers a genuine first-mover advantage within it: you would work on a unique, unpublished dataset of paired RHEED videos and XRD spectra to build one of the first video-to-spectrum models with calibrated uncertainty for oxide growth. By learning directly from experimental 1 data, such a model can short-circuit the traditional trial-and-error loop, accelerate discovery and drastically cut experimental time, cost, precursor consumption and instrument occupancy - turning routine in situ diagnostics into predictive tools.

Scientific context. Epitaxial perovskite-oxide thin films exhibit rich, tunable functional properties (thermoelectricity, ferroelectricity, piezoelectricity) governed by their structure and composition, themselves correlated with the growth conditions. MBE offers independent control of each element but, applied to oxides, is notoriously unstable and poorly reproducible because of the oxidising atmosphere. RHEED tracks the surface dynamics in real time (lattice parameter, surface reconstructions, roughness) but does not give access to the final bulk microstructure, which is only reachable ex situ via XRD / XRR. There is therefore a strong physical link - so far unexploited quantitatively - between RHEED dynamics and final XRD microstructure, which OXYD-IA proposes to learn directly.

Figure 1 - OXYD-IA pipeline. At inference, the RHEED video acquired during growth (INL) feeds a deep video encoder (LIRIS) and a probabilistic regression module that predicts the final XRD spectrum together with its uncertainty. During training, the predicted spectrum is compared with the ex situ XRD / XRR spectrum measured at INL (ground truth); this discrepancy is the supervised learning signal that adjusts the model parameters.

KEY RESEARCH AREAS AND RESPONSIBILITIES

You will drive the core computational research along a risk-managed, phased 12-month roadmap. The primary target is a versioned RHEED / XRD dataset, growth-regime classification and probabilistic regression of the key structural parameters with calibrated uncertainty; full-spectrum prediction, a real-time demonstrator and cross-oxide generalisation are stretch and follow-on objectives. The work is organised around three technical pillars:

1. Corpus structuring and spatio-temporal RHEED representation

• Build a versioned dataset by cleaning, temporally aligning and pairing the RHEED video sequences, the associated MBE parameters (temperature, oxygen pressure, source fluxes) and the ex situ XRD / XRR spectra (SrTiO3 reference system).
• Pre-train, in a self-supervised manner, a spatio-temporal video encoder (3D CNN, e.g. R(2+1)D, or temporal Vision Transformer) fine-tuned on the laboratory corpus.
• Perform fuzzy / Bayesian classification of the growth regimes (2D vs. 3D modes, transitions, RHEED oscillations), building on LIRIS expertise in the statistical analysis of sequences.

2. Probabilistic RHEED →XRD regression with uncertainty quantification

• Design a latent representation that couples the video encoding with the physical MBE parameters.
• Learn stepwise – first scalar structural parameters (lattice parameter, density, mosaicity), more robust at small N, then the full 1D XRD spectrum.
• Provide calibrated uncertainty estimates (Bayesian approaches or deep ensembles), indispensable for guiding real-time decisions during growth.

3. Real-time demonstrator and dissemination (stretch / follow-on)

• Validate the model on new SrTiO3 growths and prototype an in-growth inference tool at the end of the NANOFUTUR MBE line.
• Mitigate the small-data and generalisation challenges through temporal data augmentation, transfer learning from pre-trained video encoders and domain adaptation towards other functional oxides.
• Contribute toward a high-impact joint publication at the AI / materials interface (the project milestone is the submission of a joint paper).

REQUIRED QUALIFICATIONS

• A Ph.D. in the field of Artificial Intelligence – Computer Science, Machine Learning, Computer Vision or Signal Processing, and we are looking for a senior postdoc with 2-5 years of post-PhD experience.
• Deep learning on sequential / video data: demonstrable, hands-on experience with 3D CNNs, temporal transformers and / or probabilistic models. Proficiency in Python and PyTorch (or TensorFlow) is mandatory.
• A genuine appetite for physics: a strong curiosity for, or prior exposure to, experimental physics and materials science. The ideal candidate carried out an AI-focused Ph.D. and is eager to apply it to a real materials-physics problem.
• Research mindset: a track record of innovation and the ability to work autonomously on a complex, multidisciplinary project.
• Software skills: experience with Git version control and reproducible research.
• Excellent communication and scientific-writing skills in English.

ADDITIONAL SKILLS AND COMPETENCIES

• Self-supervised video pre-training, uncertainty quantification (Bayesian / ensembles), fuzzy classification, or domain adaptation.
• Experience with small-sample / transfer-learning settings and handling large-scale experimental datasets.