Fast Solvers for Charge Distribution Models on Shared Memory Platforms

Recent Advances for Improving the Accuracy, Transferability, and Efficiency of Reactive Force Fields

“…For this reason, they next investigated sparse approximate inverse (SAI) preconditioning techniques, as those can easily be scaled to large systems. 20 SAI preocnditioners are generally less effective than ILU techniques. However, they were able to identify sparsity patterns in SAI that allow QEq, EEM and ACKS2 solvers to convergence at about the same rate as ILU based techniques.…”

“…A recent study compared the performance of several preconditioners on the convergence of EEM and demonstrated that a significant reduction in computational time can be achieved by applying incomplete factorization and sparse approximate inverse (SAI) preconditioners. 20 Another recent addition is the development of extended Lagrangian schemes that have been demonstrated to reduce the computational cost of EEM significantly. The inertial extended Lagrangian/self-consistent scheme method (iEL-SCF) reduces the number of SCF iterations by allowing sufficient energy conservation with a high convergence tolerance.…”

“…SAI preconditioning significantly improves the convergence rate of QEq compared to Jacobi preconditioning, yielding about 3.5 and 6 times faster convergence rates at 10 −6 tolerance for silica and water, respectively. The margin between SAI and Jacobi preconditioning gets larger as we go to 10 based solver for ACKS2 in, 20 unfortunately GMRES is not a nicely parallelizable solver, and CG does not work for ACKS2 due to the indefinite nature of the matrices. However, it is possible to extend these preconditioning techniques to the ACKS2 model using a biconjugate gradient stabilized (BiCGStab) solver.…”

“…In addition, while ReaxFF offers an affordable method with respect to quantum chemistry, it is still substantially more expensive than non-reactive force fields and therefore increasing its efficiency is highly desirable. [19][20][21] This computational cost arises in part from the EEM methods that require solving a set of linear equations self consistently, in which previous work has shown that EEM can account for ∼80% of the computational time for ReaxFF applied to large systems and tight convergence tolerance. 21,22 Here we consider recent advancements that have been developed in order to address some of these issues.…”

Stochastic Constrained Extended System Dynamics for Solving Charge Equilibration Models

“…The number of self-consistent field (SCF) iterations can be reduced with careful preconditioning, polynomial extrapolation from previous steps, and good software implementations, − but solving new charge distributions at each time step remains the most computationally demanding component of MD simulations using ReaxFF, which is often 1 order of magnitude slower than traditional nonreactive force fields. Recently we have also been aware of several optimizations , in terms of preconditioners such as the sparse approximate inverse (SAI) preconditioner, as well as communication overheads through multicore architectures. While these optimizations can reduce the iteration number to 5–20 (10 –10 criteria), they may require prior knowledge about simulation systems to configure the ad hoc optimization techniques.…”