Sékou-Oumar Kaba

Ph.D. Candidate, Computer Science
McGill University & Mila – Quebec AI Institute
Montréal, Canada

Email: sekou.oumar.kaba [at] gmail.com | Scholar | Twitter | LinkedIn | GitHub

My name is Sékou-Oumar Kaba (he/him, most people call me Oumar). I am a Ph.D. candidate in Computer Science at McGill University and Mila, supervised by Prof. Siamak Ravanbakhsh.

I work on geometric deep learning and generative models for scientific discovery, with a focus on symmetry, equivariance, and AI for materials and molecular systems. Some of my recent projects span crystal generation, symmetry breaking with equivariant networks, energy-based losses for physical systems and equivariant adaptation of large pretrained models. My interests also include graph learning and statistical physics.

I previously did an internship at Microsoft Research Amsterdam working on machine learning for electronic structure. Before starting my Ph.D., I completed an M.Sc. in theoretical condensed matter physics working with Prof. David Sénéchal and did an internship in Prof. Yoshua Bengio's group at Mila.

[expand]

[expand all]

[collapse all]

Selected Publications

Conference papers

Improving Equivariant Networks with Probabilistic Symmetry Breaking (ICLR 2025). arXiv

SymmCD: Symmetry-preserving Crystal Generation with Diffusion Models (ICLR 2025). arXiv

On the Identifiability of Causal Abstractions (AISTATS 2025). arXiv

Equivariant Adaptation of Large Pre-trained Models (NeurIPS 2023). arXiv

Equivariance with Learned Canonicalization Functions (ICML 2023). arXiv

Equivariant Networks for Crystal Structures (NeurIPS 2022). arXiv

Journal articles

Prediction of Large Magnetic Moment Materials with Graph Neural Networks and Random Forests, Phys. Rev. Materials, 2023. journal

Group-theoretical Classification of Superconducting States of Strontium Ruthenate, Phys. Rev. B, 2019. journal

Preprints

Accurate and Scalable Exchange-Correlation with Deep Learning, 2025. arXiv

Full list on Google Scholar

Teaching

COMP 551 — Applied Machine Learning (Fall 2025)

Co-instructor with Prof. Reihaneh Rabbany.

Course website:Syllabus, and materials will be posted here as they become available. Assignments and quizzes are available on MyCourses.

Software

EquiAdapt — Equivariant adaptation of neural networks. docs
Equivariant Networks for Crystal Structures — reference implementations. code

Short bio

I am a Ph.D. candidate in Computer Science at McGill and Mila, supervised by Prof. Siamak Ravanbakhsh. My research focuses on geometric deep learning and generative modeling for molecules, crystals, and materials. I came to this field from physics, having completed an M.Sc. at Université de Sherbrooke and a B.Sc. at Université Laval. Along the way, I interned at Mila with Yoshua Bengio and at Microsoft Research Amsterdam, working on machine learning for materials and quantum chemistry. Beyond research, I care deeply about science outreach and communication. I have for example hosted a radio show on scientific and social innovation, and I continue to seek opportunities to make scientific ideas accessible and inspiring to a broader audience.

Link to CV (PDF)