A self-learning artificial intelligence
system for an autonomous
molecular search was recently utilized in place of laborious material
development processes by humans. In this approach, because the evaluation
of unsuitable or unrealistic candidates considerably decreases the
search efficiency, prior knowledge of the chemistry and engineering
requirements should be embedded into the molecular-generative algorithm.
However, when using naive rule-based restrictions, one must implement
the complex rule logic into the code each time, depending on the materials
and potential applications. Herein, we propose a molecular-generative
method using a maze game to control the allowable constituent fragments
of molecules, which improves the flexibility and consistency to implement
the rules. We performed an autonomous search for optimized cation
structures of high Li-ion conductive ionic liquids evaluated by molecular
dynamics simulations, in its practically reasonable scope defined
by the maze game. From the search, we discover that acyl ammonium
cations are favorable for high Li-ion conductivity because of the
high association between the cations and Li ions. These results broaden
our existing insight owing to the ability to explore beyond our practical
experiences.