MatRIS: Toward Reliable and Efficient Pretrained Machine Learning Interatomic Potentials

Paper | Poster

Official PyTorch implementation of MatRIS (Materials Representation and Interaction Simulation), a pretrained interatomic potential for materials simulation.

Update

Oct 2026

• Optimized kernels and torch.compile support (~5× faster inference).
• TorchSim integration.
• New models: OMat24 and the MatPES PBE/r2SCAN family.

Installation

• Python ≥ 3.12, PyTorch ≥ 2.11
• NumPy, ASE, pymatgen, nvalchemi-toolkit-ops, Ninja

Pretrained Models

We offer seven pretrained models:

Model key	Checkpoint	Dataset	Target
matris_10m_omat	Download	OMat24	efs
matris_10m_oam	Download	OMat24 → sAlex + MPtrj	efsm
matris_10m_mp	Download	MPtrj	efsm
matris_4m_matpes_pbev1	Download	MatPES-PBE (2025.1)	efs
matris_4m_matpes_pbev2	Download	MatPES-PBE (2025.2)	efs
matris_4m_matpes_r2scanv1	Download	MatPES-r2SCAN (2025.1)	efs
matris_4m_matpes_r2scanv2	Download	MatPES-r2SCAN (2025.2)	efs

If you need other models, feel free to contact me.

Usage

There are some examples how to use MatRIS, including calculator, geometry optimization, molecular dynamics and TorchSim.

ASE Calculator

from ase.build import bulk
import torch

from matris.applications.base import MatRISCalculator

device = "cuda" if torch.cuda.is_available() else "cpu"
calc = MatRISCalculator(
    model="matris_10m_oam", # model name or checkpoint path
    task="efsm",
    device=device,
    mode="fast",                 # optimized kernel + compile; "torch" for eager mode
    activation_checkpoint=True,  # set True for adaptive memory-saving recomputation
    tf32=False,                  # set True to allow TF32 in interaction blocks (~1.5x faster)
)

atoms = bulk("Cu", cubic=True)
atoms.calc = calc

energy = atoms.get_potential_energy()   # total energy, eV
forces = atoms.get_forces()             # eV/A
stress = atoms.get_stress()             # eV/A^3, ASE convention
magmoms = atoms.get_magnetic_moments()  # muB

Structure Optimization

from ase.build import bulk
import torch

from matris.applications.relax import StructOptimizer

device = "cuda" if torch.cuda.is_available() else "cpu"
optimizer = StructOptimizer(
    model="matris_10m_oam",
    task="efsm",
    optimizer="FIRE",
    device=device,
    mode="fast",
    activation_checkpoint=False,
    tf32=False,
)

atoms = bulk("Cu", cubic=True)
result = optimizer.relax(
    atoms=atoms,
    verbose=True,
    steps=500,
    fmax=0.05,
    relax_cell=True,
    ase_filter="FrechetCellFilter",
)

final_structure = result["final_structure"]
trajectory = result["trajectory"]

Molecular Dynamics

from ase.build import bulk

from matris.applications import MolecularDynamics

atoms = bulk("Cu", cubic=True)

md = MolecularDynamics(
    atoms=atoms,
    model="matris_10m_oam",
    ensemble="nvt",
    temperature=300,
    starting_temperature=300,
    timestep=1,
    trajectory="md_out.traj",
    logfile="md_out.log",
    loginterval=100,
    task="efsm",
    device="cuda",
    mode="fast",
    activation_checkpoint=False,
    tf32=True,
)
md.run(1000)

md.set_atoms(new_atoms) restarts the integrator and step count for a new system, retains the calculator, and appends trajectory output. Reattach custom ASE observers to md.dyn after replacing atoms.

TorchSim

Install torch-sim-atomistic, then use pretrained names or local weights:

import torch
import torch_sim as ts
from ase.build import bulk
from matris.applications.torchsim import MatRISTorchSimModel

model = MatRISTorchSimModel(
    model="matris_10m_oam",
    target="efsm",  # OAM/MP: efsm; OMat/MatPES: efs
    device="cuda",
    mode="fast",
    activation_checkpoint=True,
    tf32=False,
)
state = ts.io.atoms_to_state(
    [bulk("Si", "diamond", a=5.43)], device=model.device, dtype=torch.float32,
)
results = model(state)  # energy (eV), forces (eV/Å), stress (eV/ų), magmoms (μB)
state = ts.integrate(
    state, model, integrator=ts.Integrator.nvt_langevin,
    n_steps=100, temperature=300, timestep=0.001,  # ps
)

Citation

If you use MatRIS in your work, please cite:

@inproceedings{
zhou2026matris,
title={Mat{RIS}: Toward Reliable and Efficient Pretrained Machine Learning Interatomic Potentials},
author={Yuanchang Zhou and Siyu Hu and Xiangyu Zhang and Hongyu Wang and Guangming Tan and Weile Jia},
booktitle={The Fourteenth International Conference on Learning Representations},
year={2026},
url={https://openreview.net/forum?id=5xBT5Ziute}
}

License

MatRIS is licensed under the BSD-3-Clause License. See LICENSE for details.