MLIP Dynamics

Temperature–

Energy drift–

Speed–

Compute–

Zero-point energy–

Walkers–

Throughput–

Compute–

Energy, eV, relative to the total at t = 0

Temperature, K

Reference energy E_ref − V_min, cm⁻¹

Walkers

Velocity Verlet at constant N, V and E. When the forces are the gradient of the energy and Δt is small enough, the total energy stays flat while kinetic and potential energy swap. From a relaxed start about half the initial kinetic energy goes into the potential, so the temperature settles near T₀/2. Direct forces are not a gradient, so the total energy drifts with them, and so it does with too large a Δt.

Diffusion Monte Carlo for the nuclei. The structure is relaxed on the model's surface, then each walker, a copy of the molecule, takes Gaussian steps of variance ħ²Δτ/m and is copied or removed with weight exp(−Δτ (V − E_ref)). E_ref is adjusted to keep the number of walkers near the target, and its average above the minimum is the zero-point energy. The cloud shows the walkers aligned on the minimum; hydrogens spread the most. Walkers far below the minimum are in a hole in the model's surface and are removed.

FLOPs are counted from the shapes of the kernels the model runs, with a multiply-add as 2. The rate is over the whole step, so it also covers the neighbour list, data transfers and reading back the results.