Targeted Adversarial Learning Optimized Sampling

Boosting transitions of rare events is critical to modern-day simulations of complex dynamic systems. We present a novel approach to modify the potential energy surface in order to drive the system to a user-defined target distribution where the free energy barrier is lowered. The new approach, called targeted adversarial learning optimized sampling (TALOS), cross-fertilizes statistical mechanics and deep learning. By casting the enhanced sampling problem as a competing game between a real sampling engine and a virtual discriminator, TALOS enables unsupervised construction of bias potential on an arbitrary dimensional space and seeks for an optimal transport plan that transforms the system into target. Through multiple experiments we show that on-the-fly training of TALOS benefits from the state-of-art optimization techniques in deep learning, thus is efficient, robust and interpretable. TALOS can also simultaneously learn to extract good reaction coordinate from a high-dimensional space where bias potential is being constructed. Additionally, TALOS is shown to be closely related to reinforcement learning, giving rise to a new framework of manipulating Hamiltonian in order to fulfill user-specified tasks via deep learning.