Anirban Chandra1,Rohit Batra2,Amanda Dufek3,Suvo Banik1,Isaac Tamblyn4,Pierre Darancet2,Stephen Whitelam3,Subramanian Sankaranarayanan2,1
University of Illinois at Chicago1,Argonne National Laboratory2,Lawrence Berkeley National Laboratory3,University of Ottawa4
Anirban Chandra1,Rohit Batra2,Amanda Dufek3,Suvo Banik1,Isaac Tamblyn4,Pierre Darancet2,Stephen Whitelam3,Subramanian Sankaranarayanan2,1
University of Illinois at Chicago1,Argonne National Laboratory2,Lawrence Berkeley National Laboratory3,University of Ottawa4
Discovering pathways to efficiently transition between metastable and stable phases is important from both a computational and experimental standpoint. Nucleation of crystalline phases from disordered phases is typically a slow process and can be challenging to observe in timescales accessible by molecular simulations. These in turn often impact polymorph selection - this is important given that the properties are significantly different depending on the polymorph type. Here, using the Learning to grow workflow developed by Whitelam and Tamblyn, we investigate how neural networks trained by evolutionary reinforcement learning can generate thermodynamic pathways that facilitate nucleation in subcooled water and enable us to attain a targeted polymorph. Using the same methodology we explore nucleation of a single crystal hexagonal ice, the most stable ice polymorph but attaining which has been a challenging task for molecular simulations till date. Our neural networks, after being trained on molecular simulation trajectories, learn to change thermodynamic variables in order to promote the generation of a particular phase. These previously unknown pathways can enable us to optimally create a particular material and at the same time, provide physical insight into phase transition kinetics. The presented evolutionary scheme can also be applied to a broad range of problems (such as self-assembly) wherein transition pathways from one state to another are unknown.