Reinforcement learning for dynamic control of self-assembly of quasicrystals from patchy particles

Patchy particles are the colloidal particles exhibiting anisotropic interactions with other particles. These patchy particles are capable of organising into complex structures, e.g. quasicrystalline structures, which are important to the development of novel materials. However, controlling the assembly of these particles is challenging because the structures at the steady states are significantly influenced by the kinetic pathways of their structural formation. It is well known that the steady-state structure is largely affected by dynamic control factors such as the change in temperature and external mechanical forces. Reinforcement learning is a branch of machine learning that aims to acquire an optimal policy and protocol for interacting with the environment through experience. Reinforcement learning can estimate an external force or parameter change as a function of the state of the system. In this study, we propose using reinforcement learning to control the dynamical self-assembly of the dodecagonal quasicrystals from patchy particles. We estimate the best policy of temperature control trained through the Q-learning method and demonstrate that we can generate dodecagonal quasicrystals with few defects by applying the estimated policy.