Project A10 (New): Population control of multiple walker simulations via a birth/death process
Conventional Molecular Dynamics (MD) simulations are generally unable to access the long-timescale phenomena that are common in nature. This timescale problem comes from the fact that a typical free energy landscape consists of many metastable states separated by high free energy barriers. If the barriers are much higher than the thermal energy, the system is kinetically trapped in some metastable state and barrier crossings will be rare events on the time scales that we can simulate. One strategy to alleviate this time scale problem is to employ collective variable (CV) based enhanced sampling methods such as metadynamics. A common way to improve the performance of CV-based methods is to employ multiple walkers that share a bias potential and collaboratively sample the free energy landscape. In this way, one reduces the wall-clock time for convergence and makes better usage of modern parallel HPC resources. However, the multiple walkers’ usefulness can be severely diminished due to walkers becoming correlated with each other. Thus, there is a need to develop a population control scheme for walkers to improve performance. In this project, we propose to develop a population control scheme for multiple walker CV-based enhanced sampling based on a recently introduced birth/death accelerated Langevin sampling algorithm (Lu et al., arXiv preprint at arXiv:1905.09863 ). We will adapt this birth/death scheme to atomistic/molecular systems and integrate it into a CV-based multiple walker framework.
The birth/death algorithm in the paper by Lu et al. is based on a Fokker-Planck equation with an additional birth/death term. The advantage of the birth/death term is to allow for global moves of the probability density mass between metastable modes without having to go through regions of low probability (i.e., barriers). In practice, the algorithm is implemented by considering an ensemble of walkers. Between birth/death events, each walker independently follows Langevin dynamics in the given energy landscape. Then periodically, one kills or duplicates walkers according to their birth/death rate. The non-local birth/death moves can significantly improve the sampling of the energy landscape. However, this scheme has some shortfalls. It cannot describe higher-lying free energy regions as walkers there tend to be killed. Furthermore, one must make sure that the initial distribution of walkers covers all the metastable states as the scheme cannot explore and find new metastable states on its own if the free energy barriers are too high. In other words, the method is based on the exploitation of information. On the other hand, CV-based methods are based on exploration and can find
new metastable states on their own. Therefore, combining the birth/death scheme with multiple walker CV-based enhanced sampling can improve both methods' performance.
The development of this birth/death scheme will require overcoming several theoretical and practical challenges. This project will thus be performed in a tight collaboration between theoretical physics (Burkhard Dünweg and Omar Valsson) and mathematics (Lisa Hartung).
Improving the Efficiency of Variationally Enhanced Sampling with Wavelet-Based Bias Potentials
Benjamin Pampel and Omar Valsson
Journal of Chemical Theory and Computation 2022, 18, 7, 4127–4141