Rees Chang1,Nick Richardson2,Alex Guerra2,Angela Pak1,Ni Zhan2,Alex Ganose3,Adji Dieng2,Ryan Adams2,Elif Ertekin1
University of Illinois at Urbana-Champaign1,Princeton University2,Imperial College London3
Rees Chang1,Nick Richardson2,Alex Guerra2,Angela Pak1,Ni Zhan2,Alex Ganose3,Adji Dieng2,Ryan Adams2,Elif Ertekin1
University of Illinois at Urbana-Champaign1,Princeton University2,Imperial College London3
This presentation will highlight our explorations of approaches to deep generative models for atomic crystals. Generative modeling of crystals is challenging due to the hard constraints with respect to symmetry and stability, variable lattice sizes and numbers of atoms, and mix of both discrete and continuous attributes. Existing methods aim to learn the distribution of stable atomic configurations without consideration of space group symmetries. To overcome this limitation, we are using highly interpretable GFlowNets to generate crystals approximately from the Boltzmann distribution and conditioned on space group and/or chemistry. The model is also explicitly trained to avoid creating invalid structures. We will present the performance and implications of our framework, highlighting the advantages of symmetry-aware generation towards inverse design of materials.