New math method speeds molecular simulations on supercomputers
Flatiron Institute researchers report a 2.5- to sevenfold acceleration for molecular dynamics simulations without loss of accuracy.
By Priya Raghavan · Science Reporter
3 min read
Researchers at the Simons Foundation’s Flatiron Institute have developed a mathematical method that can speed up molecular dynamics simulations by 2.5 to seven times, according to a study in Nature Communications. The advance could cut computing time and energy use for simulations used in materials research, drug studies and protein-folding work.
The work targets a costly part of high-performance computing. According to the Simons Foundation, more than 20% of the workload on the world’s 500 fastest supercomputers goes to modeling the motion of atoms and molecules.
The team reported a fivefold speedup in high-accuracy runs using GROMACS, a widely used molecular dynamics software package. The researchers said the method can be added to existing software workflows, including LAMMPS, GROMACS and OpenMM; Shidong Jiang of the Flatiron Institute said the LAMMPS developers have accepted the team’s code.
Why the calculations take so long
Molecular dynamics simulations follow systems such as a protein in water or an electrolyte in a battery by breaking time into tiny steps. Pilar Cossio, a Flatiron Institute senior research scientist who uses the simulations, described the task as tracking how such a system changes over time.
The required time resolution is severe. The Simons Foundation said molecular bond vibrations require about 500 trillion slices of time per second, and useful simulations may need up to a trillion steps. Sonya Hanson, another Flatiron Institute researcher who works with molecular dynamics, said even strong hardware can leave researchers advancing only hundreds of nanoseconds per day, stretching microsecond-scale studies over weeks.
A major bottleneck is the calculation of electrostatic forces among charged particles. Jiang said a direct approach would require work proportional to the square of the number of atoms, making large systems impractical without mathematical shortcuts.
Researchers already rely on tools such as the fast Fourier transform and the fast multipole method, the latter co-invented by Flatiron Institute computational mathematician Leslie Greengard. The new study aims to improve the long-range force calculations that still consume substantial computing resources.
A 19th-century function applied to a modern bottleneck
The Flatiron team used prolate spheroidal wave functions, a class of functions first introduced in 1880 and later used in signal processing at Bell Labs in the 1960s. According to the Simons Foundation, the functions help decide how to divide electrostatic interactions into short-range and long-range parts and how to place atomic charges onto a grid for long-range calculations.
Those operations need functions that are both tightly confined in space and very smooth. The researchers found that prolate spheroidal wave functions satisfy those competing requirements better than functions commonly used before, the Simons Foundation said.
The study was led by Jiuyang Liang, a Flatiron Institute affiliate research fellow and Shanghai Jiao Tong University researcher. Co-authors include Libin Lu, Alex Barnett, Jiang and Greengard, all affiliated with the Flatiron Institute’s Center for Computational Mathematics.
The team tested the method on several systems, including water molecules, an immune-related protein system and a lithium-ion solution used in battery research. Across those tests, the researchers reported speedups ranging from 2.5 to seven times while maintaining accuracy.
Anthony Costa, director of digital biology at Nvidia, who was not involved in the study, said the work has strong potential to accelerate molecular dynamics workloads, a field he said has seen only incremental speed gains over the past decade.
This story draws on original reporting from Phys.org.