QuickMDSim
All posts

Molecular dynamics just got faster — and LAMMPS is getting it

QuickMDSim 5 min read

Long-range electrostatics is the silent tax on every molecular dynamics job. A team at the Flatiron Institute just found a way to cut it — and their code is being folded into LAMMPS.

The news

In a paper published in Nature Communications (May 2026), researchers at the Simons Foundation's Flatiron Institute showed a new mathematical method that runs classical molecular dynamics simulations 2.5 to 7 times faster — without sacrificing accuracy. For GROMACS at high accuracy, they measured a fivefold speedup.

The method, from lead author Jiuyang Liang and colleagues (Shidong Jiang, Alex Barnett, Leslie Greengard, and others), has been officially accepted by the LAMMPS developers. That matters to us: it means the workhorse engine we build on gets faster the moment it lands.

2.5–7×
Speedup, no accuracy loss
>20%
of top-500 supercomputer work is atomistic sim
1M
atoms in the LiTFSI battery-electrolyte test

Why long-range forces are the bottleneck

MD steps time in tiny slices — capturing molecular bond vibrations needs roughly 500 trillion timesteps per second. A meaningful simulation runs a trillion steps. Over a few microseconds of simulated time, that adds up to "hundreds of nanoseconds a day" on great hardware.

The killer is electrostatic interactions. Charged particles push and pull over long distances, so in principle you must account for every atom pair. Done naively, that's operations proportional to the square of the atom count. Decades of math — fast Fourier transforms, the fast multipole method — have made it cheaper, but long-range force calculation still eats the bulk of runtime.

The math: smoothing the speed bumps

The Flatiron team reached back to an 1880s invention, prolate spheroidal wave functions (later used for signal processing at Bell Labs in the 1960s). They use these functions to make two decisions in the simulation: how to split electrostatics into short-range and long-range parts, and how to spread atomic charges onto a grid for the long-range part.

The functions need to be both spatially localized and as smooth as possible — a pair of demands older methods compromised on. The prolate function satisfies both better, so the sharply spiked electric potential gets smoothed into something far cheaper to compute. Fewer pairwise sums, same physics.

"There are so many fields in science that rely on molecular simulations that can now take less energy and computing time." — Shidong Jiang, Flatiron Institute

What we ran in QuickMDSim

To make this concrete, we ran a small crystal-melt job in QuickMDSim — a Lennard-Jones system heated through a phase transition. The screenshot below is the simulation output: an ordered FCC lattice seeding thermal defects and finally melting into a disordered liquid.

A molecular dynamics crystal-melt render from QuickMDSim: an ordered amber, sky, and violet FCC lattice seeding thermal defects and melting into a disordered liquid.
A crystal-melt trajectory rendered from a QuickMDSim job. Warm colors intensify as the lattice heats and disordering begins.

It's a deliberately small system — the point isn't the scale, it's that the workflow is unremarkable: create a project, paste the input, submit, download. No cluster module to load, no scheduler to wait on. That's the whole idea.

Why this matters for how we run MD

Efficiency gains in MD don't stay in the lab that found them — they compound across every practitioner. When the new Ewald method lands in LAMMPS, every LAMMPS run inherits it: shorter queues, cheaper jobs, more science per compute-hour.

That's exactly the problem QuickMDSim exists to attack. We're building a workbench where the infrastructure is out of the way, and the engine keeps getting faster underneath you. Cloud LAMMPS without the cluster tax — the tax shrinking from both directions.

Curious? Open the app and run a first job in a few minutes. Or email us — we read everything.