Kinetically Controlled Self-assembly via Reinforcement Learning

Controlling self-assembly is challenging when multiple competing products (i.e., polymorphism) exist. In this work, we study self-assembly of a model patchy particle with polymorphic assembly and show how the final product can be controlled by an adaptive policy learned using the deep Q-network (DQN) method. Conventional molecular dynamics (MD) and Hamiltonian replica exchange (HREX) MD simulations show that these particles self-assemble into two morphologies - rhombohedron and trihexagonal crystalline structures. The distributions of the Steinhardt order parameter-, q₆ obtained from conventional and HREX MD suggest that the trihexagonal and rhombohedron structures are thermodynamically and kinetically stable, respectively. The HREX simulations also suggest that the fraction of each morphology produced can by altered by changing the assembly temperature. Inspired by our preliminary results, we employ the DQN method to learn a temperature-based protocol to preferentially obtain the kinetically controlled state. The DQN method employed here provides an effective way of controlling self-assembly, and can be extended to other systems – for example, in biological assemblies.