Efficient and Accurate Spatial Mixing of Machine Learned Interatomic Potentials for Materials Science

TL;DR

ML-MIX addresses the cost-accuracy trade-off of ML interatomic potentials by spatially mixing a cheap linear potential with expensive MLIPs via a QM/MM–inspired force-mixing framework integrated into LAMMPS. It distills cheap potentials through constrained fitting and region tracking to locally approximate the expensive potential, enabling large-scale, CPU- and GPU-accelerated simulations with minimal loss of accuracy. Across Si, Fe, and W, including defects, diffusion, screw dislocation dynamics, and He implantation, ML-MIX achieves speedups up to about 11× for systems of ~8,000 atoms while reproducing key quantities and even matching experimental reflection coefficients up to 80 eV in W–He implantation. The work demonstrates the practical deployment of state-of-the-art MLIPs on realistic, large-scale systems and outlines limitations (notably energy non-conservation with force-mixing) and future directions (uncertainty quantification, broader MLIP support, and energy-mixing strategies).

Abstract

Machine-learned interatomic potentials can offer near first-principles accuracy but are computationally expensive, limiting their application to large-scale molecular dynamics simulations. Inspired by quantum mechanics/molecular mechanics methods we present ML-MIX, a CPU- and GPU-compatible LAMMPS package to accelerate simulations by spatially mixing interatomic potentials of different complexities allowing deployment of modern MLIPs even under restricted computational budgets. We demonstrate our method for ACE, UF3, SNAP and MACE potential architectures and demonstrate how linear 'cheap' potentials can be distilled from a given 'expensive' potential, allowing close matching in relevant regions of configuration space. The functionality of ML-MIX is demonstrated through tests on point defects in Si, Fe and W-He, in which speedups of up to 11x over ~ 8,000 atoms are demonstrated, without sacrificing accuracy. The scientific potential of ML-MIX is demonstrated via two case studies in W, measuring the mobility of b = 1/2 111 screw dislocations with ACE/ACE mixing and the implantation of He with MACE/SNAP mixing. The latter returns He reflection coefficients which (for the first time) match experimental observations up to an He incident energy of 80 eV - demonstrating the benefits of deploying state-of-the-art models on large, realistic systems.

Figures (12)

• Figure 1: A schematic overview of the simulation acceleration workflow. Left side: the constrained potential fitting process, whereby an accurate expensive potential (orange) is approximated in local regions of potential energy space by a simpler cheap potential (blue). Right side: schematic showing how simulation is accelerated; cheap potential is used to evaluate force components on atoms in 'bulk-like' environments.
• Figure 2: The energy barrier associated with vacancy migration in Si, obtained through all-expensive (blue curve) and ML/ML (orange curve) nudged elastic band simulations. For the ML/ML simulation, the final energies were obtained through one-shot energy evaluations of the relaxed structures with the expensive potential. The all-expensive energy barrier (0.0543 eV) agrees with the ML/ML energy barrier (0.0548 eV) within 1 meV. The tangent lines arise from the projected forces.
• Figure 3: Average force on stretched bond in silicon over 100 ps at 300 K for different expensive potential radii $r_{\mathrm{core}}$ (blue points) . Error bars represent the standard error in bond force measured over this time, with samples taken every 15 fs. For $r_{\mathrm{core}} = 0$, which corresponds to using the cheap potential everywhere, the average bond force measured does not match the all-expensive $NVE$ reference (red dashed line) . Once the expensive potential is introduced at and around the stretched bond, this difference vanishes and agreement between the mixed simulation and the $NVE$ reference is within statistical error.
• Figure 4: Diffusion coefficient as a function of temperature for the diffusion of the iron interstitial dumbbell defect through bulk iron and for helium in bulk tungsten. (a) : The diffusion coefficient of the iron dumbbell for three cases: expensive reference potential only (blue), cheap potential only (orange) and mixed simulation (green). The mixed simulation leads to results which are well matched with the reference. Each point is the average of 5 diffusion coefficients measured in independent runs of 1 ns each. (b) : The diffusion coefficient for He in W using a helium-tungsten ACE (blue) and a mixed simulation with both a helium-tungsten ACE and a UF3 tungsten potential (orange). Again, agreement is good. Note that there is no 'cheap-only' reference in this case because the UF3 potential cannot be used to model tungsten-helium interactions. Each point is the average of 25 diffusion coefficients measured in independent runs of 60 ps each. The error-bars on the points in both (a) and (b) represent the standard error in each value.
• Figure 5: Simulation set-up and results for thermally activated glide of screw dislocations in W at different shear stresses and temperatures. (a) Simulation set-up, where dimensions in $x$, $y$ and $z$ are $223~\mathrm{\AA} \times 221~\mathrm{\AA} \times 68~\mathrm{\AA}\ (25~\textit{b})$, respectively, for a total of 195,000 atoms. Stress was applied to the free surfaces at the top and bottom of the cell in $\langle111\rangle$ type directions. The expensive atoms around the dislocation (as detected using common neighbor analysis in LAMMPS) are shown in a column, where the colors correspond to the core (red), blending (green) and buffer (white) regions. (b) The velocity results at a range of temperatures and shear stresses. Green, orange and blue lines represent 300, 600 and 900 K respectively. Results from all-expensive simulations are shown with dashed lines, whilst those from mixed simulations are shown with solid lines. For each point, three separate simulations were conducted and the results were averaged. Results from the mixed simulations agree well with those obtained from the all-expensive reference. The error-bars on the points represent the standard error in each value.