I developed a graph-based research framework in Python and PyTorch for molecular structure elucidation using discrete diffusion models and stochastic interpolants.

The framework supports Transformer architectures, custom sampling strategies, and guidance mechanisms. It serves as a platform for developing and evaluating new generative methods for molecular structure prediction. The models were trained and evaluated on the Jean Zay national supercomputer.

This is joint work with Jérémie Cabessa, Thibault Charpentier, Marie-Pierre Gaigeot, and Mihai-Cosmin Marinica.

Technologies: Python, PyTorch, Lightning, Weights & Biases, and Slurm.
Status: The repository is currently private. A public release is planned alongside the associated publication; access may be provided on request.

I contributed to DFTK, a Julia package for plane-wave density-functional theory calculations.

My contributions include:

• an interface to Wannier90 for Wannierization, which later supported the integration of Wannier.jl into DFTK;
• an implementation of the modified-operator method.

Technologies: Julia, DFTK, Wannier90, and plane-wave density-functional theory.

With Étienne Polack, I developed a Julia package for computing band diagrams of twisted bilayer graphene.

The package implements the Bistritzer–MacDonald model and the model introduced by Cancès, Garrigue, and Gontier, Physical Review B 107, 2023. It is built on top of DFTK.

Further mathematical and numerical details are available in Chapter 5 of my PhD thesis.

Technologies: Julia, DFTK, and numerical linear algebra.
Status: Research implementation; not currently publicly released.

• Scientific machine learning: Python, PyTorch, Lightning, and Weights & Biases; implementation, training, and evaluation of diffusion and Transformer-based architectures.
• Scientific computing: Julia and DFTK, with applications to electronic-structure calculations and numerical methods. Previous experience with Fortran 90 and C++. Additional contributions are available on GitHub.
• High-performance computing: Slurm and extensive use of the Jean Zay national supercomputer.