(169df) Efficiently Screening Metal-Organic Frameworks Via Molecular Simulation with Multi-Armed Bandit Algorithms

Metal-organic frameworks (MOF) are nanoporous materials composed of metal ions/clusters and organic ligands. They have large specific surface areas and a high density of functional groups. These features allow MOFs to store, sense, separate, and controlled-release gases. The porosity, pore size and shape, and chemistry of MOFs are highly adjustable. Among millions of conceivable MOFs for a given adsorption task, we can screen them using molecular simulations. However, brute-force screening is often infeasible in terms of cost and time.

This work applies multi-armed bandit optimization (MABO) to leverage short, cheap, and noisy molecular simulations to screen a large set of MOFs while incurring the least computational cost. Each MOF is analogous to a slot machine; conducting a Monte Carlo molecular simulation is analogous to pulling an arm of a slot machine; and the noisy, estimated property of a MOF from a simulation is analogous to receiving a stochastic reward from a slot machine. MABO adaptively selects a sequence of MOFs for cheap, noisy simulations while learning from feedback and balancing exploration and exploitation in its decision-making. After each simulation, MABO uses the noisy estimate of this MOF's property to update a posterior distribution of that MOF's property. Then, MABO selects the next MOF for a simulation based on the posterior distributions of the property of all candidate MOFs. Its decision balances the yearning to select a MOF with (i) a high mean property (exploitation) and (ii) a high variance (i.e., uncertainty) in the property (exploration).