Automating Simulations of Block Copolymers to Find Structural Features Using a Closed-Loop Optimization Process

When performing molecular dynamics (MD) simulations with adjustable parameters, one may consider a grid of equally spaced values of the parameters to understand the entire parameter space. However, if the goal is to find a certain region of the space (e.g. conditions that optimize a property), the user may instead search by iteratively running a small number of simulations and then choosing where to continue the search based on knowledge of the system thus far. To increase the efficiency of such a process, we automate our simulations and apply a Bayesian Optimization (BO) algorithm to choose parameters in an unsupervised manner. We test our process by finding the coil-to-globule transition of amphiphilic block copolymer chains as a function of hydrophobic fraction and solvent quality. Specifically, we use a Dissipative Particle Dynamics (DPD) model bead-spring chain in solvent with a variable solvent-bead interaction strength. A few initial simulation runs are used to generate a coarse approximation of the radius of gyration function, then a manifold-crawling BO algorithm is applied to choose additional points to determine the extremum in its derivative. We will discuss how this coil-to-globule transition depends on hydrophobic fraction for di-, tri-, and tetra-block copolymers.