Posted: 21st April 2026

DPhil student Emily Jin has co-authored a new paper introducing OXtal, a generative model for molecular crystal structure prediction that marks a significant step forward for materials and drug discovery. 

For decades, predicting how molecules pack into solids has been regarded as a grand challenge in chemistry. Crystal packing dictates the properties of solid-state materials, from the stability of pharmaceutical compounds to the performance of flexible transistors. Traditional computational approaches rely on vast brute-force searches combined with quantum chemistry calculations, often requiring millions of CPU hours. 

OXtal takes a fundamentally different approach. Rather than ‘shaking atoms’ through exhaustive search, the model directly generates experimentally realistic molecular crystal packings from a 2D molecular graph in seconds. It learns how a molecule behaves, and how multiple copies arrange themselves in a periodic crystal. This dramatic increase in speed and efficiency makes high-quality structure prediction far more accessible, with important implications for accelerating materials design and drug development. 

What excites me most about this work is that it really marks a shift in mindset for a problem that has existed for decades. From the chemistry side, OXtal moves crystal structure prediction from a brute-force search problem to a learning problem. And from the machine learning side, changing how we choose to represent the data has allowed us to unlock a whole new level of performance. Watching local chemical 'rules' give rise to global crystal order feels like a glimpse into a new design paradigm that still has much more potential for exploration. The interdisciplinary nature of this work is also a testament to all the wonderful collaborators who came together to make this project possible. Emily Jin

Emily co-authored the paper with Cheng-Hao Liu and Frances H. Arnold from Caltech, Andrei Nica from Synteny, Professor Michael Bronstein and former Research Associate Joey Bose at Oxford (Joey is now faculty at Imperial College London), and other researchers at AITHYRA, Mila, Université de Montréal, Google, Lila, and NVIDIA. They will be presenting the paper in Brazil on 24 April at the International Conference on Learning Representations 2026. 
 
You can find out more about OXtal by reading this blog piece or the full paper here.