Tin Hadzi Veljkovic
I’m a PhD student at the University of Amsterdam (AMLab / Bosch Delta Lab), advised by Jan-Willem van de Meent.
I work on generative models for materials, continuous representations, and representation learning for computational chemistry. I like building models and representations that stay grounded in physics, but are also usable in practice.
Selected work
Publications
-
Crystalite: A Lightweight Transformer for Efficient Crystal Modeling
Introduces a lightweight diffusion Transformer for crystal modeling, combining subatomic tokenization with geometry-aware attention biases to improve efficiency while reaching strong crystal structure prediction and generation results.
-
CORDS: Continuous Representations of Discrete Structures
Invertible mappings from variable-size sets to continuous density and feature fields, enabling models to operate in field space and decode exactly back to discrete objects.
-
Dynamic Training Enhances Machine Learning Potentials for Long-Lasting
Molecular Dynamics
Refreshes ML interatomic potentials during long MD runs to preserve accuracy over extended simulations.
-
Fast yet Safe: Early-Exiting with Risk Control
Applies distribution-free risk control to early-exit neural nets so faster predictions still satisfy target error guarantees.
Recent writing
Blogposts
-
Crystalite: A Lightweight Transformer for Efficient Crystal Modeling
An overview of Crystalite, its chemistry-aware design, and why lightweight transformer choices matter for crystal modeling.
Contact