Self-assembling of polygons via machine learning

Bayesian optimization (BO) is employed to determine the optimal parameters for a two-dimensional molecular building block designed to self-assemble into specific polygonal structures. The building block is a simplified model consisting of a central soft repulsive disk, decorated with two attractive patches near its perimeter, separated by an angle. The soft repulsion of the central disk using the Weeks-Chandler-Andersen (WCA) potential, while the patch attractions are modeled via the Lennard-Jones potential (LJ). The optimization seeks parameters that maximize an objective function defining the abundance of the target polygon while suppressing competing species. We introduce an inverse-design workflow that couples BO with molecular-dynamics (MD) simulations for polygon selection. A selectivity-driven objective function and polygon classifier guide the search toward high-yield conditions. Computationally, the BO–MD coupling identifies promising regions of parameter space with O ( 1 0 1 − 1 0 2 ) MD evaluations rather than a combinatorial scan. The workflow is transferable to other targets by redefining the objective/classifier and adjusting parameter bounds. The parameters optimized by BO are the patch separation angle, the radial distance from the disk center to the patch centers, and the cutoff radius of the attractive interactions; all other parameters (e.g., disk diameter and LJ well depth) are held constant. After optimization, canonical MD simulations are run for the optimized parameter sets, followed by density scans at fixed temperature. Results are presented as phase-diagram color maps for each target polygon as a function of its number of sides.