Dipole Moments and Polarizabilities with MACE

Using MACE-MDP: Pretrained Model for Dipole Moments and Polarizabilities

If you need ready-to-use predictions of molecular dipole moments and fully anisotropic polarizability tensors for organic systems, the pretrained MACE-MDP model is the recommended starting point. It is trained on the SPICE-α dataset (>1.8 million electric-field response calculations across molecular and condensed-phase environments) and achieves first-principles accuracy at a fraction of the cost. See the Foundation models page for the full model listing.

Warning

Fine-tuning of MACE-MDP is not available at this time.

The model is available via the mace_mdp convenience function, which automatically downloads and caches the model on first use:

from mace.calculators import mace_mdp
from ase import build

calc = mace_mdp(device="cuda", default_dtype="float64")
atoms = build.molecule("H2O")
atoms.calc = calc

Getting Dipole Moments

Use ASE’s standard get_dipole_moment() after attaching the calculator:

mu = atoms.get_dipole_moment()  # returns array of shape (3,) in e·Å
print("Dipole moment (e·Å):", mu)

Getting Polarizability

Use get_property with the "polarizability" key to obtain the full 3×3 tensor:

import numpy as np

alpha = np.asarray(calc.get_property("polarizability", atoms)).reshape(3, 3)
print("Polarizability tensor (Å³):\n", alpha)

Tip: For the spherical (irreducible) polarizability components, useful for Raman spectra, use the "polarizability_sh" property:

spherical_alpha = calc.get_property("polarizability_sh", atoms)

Tutorial Jupyter notebooks for IR spectra, Raman spectra, and dipole/polarizability extraction are provided in the examples/ directory of the MACE-MDP repository.

If you use MACE-MDP in your work, please cite [1].

Training Your Own Model

Training Example

A typical training command for an AtomicDielectric MACE model looks like:

python /../mace/mace/cli/run_train.py \
     --name="mace_mu_alpha" \
     --train_file="train.xyz" \
     --valid_file="val.xyz" \
     --test_dir="test.xyz" \
     --model="AtomicDielectricMACE" \
     --E0s="average" \
     --num_interactions=2 \
     --num_channels=128 \
     --max_L=2 \
     --correlation=3 \
     --MLP_irreps="16x0e+16x1o+16x2e" \
     --dipole_key="REF_dipole" \
     --polarizability_key="REF_polarizability" \
     --loss="dipole_polar" \
     --weight_decay=5e-10 \
     --polarizability_weight=2000 \
     --dipole_weight=1000 \
     --clip_grad=1.0 \
     --batch_size=128 \
     --valid_batch_size=128 \
     --max_num_epochs=40 \
     --scheduler_patience=15 \
     --patience=15 \
     --eval_interval=1 \
     --ema \
     --error_table="DipolePolarRMSE" \
     --default_dtype="float64" \
     --device=cuda \
     --seed=123 \
     --restart_latest \
     --save_cpu

Setting –MLP_irreps=”16x0e+16x1o+16x2e” and –max_L=2 are crutial for predicting polarizability correctly. Compared to a MACE - MLIP these models usually need less epochs to converge.

Extracting Polarizability Using ASE

Once training is complete and you have your model file (e.g., mace_mu_alpha.model), you can extract polarizability tensors from a trajectory.

Example extraction script:

import numpy as np
from ase.io import Trajectory
from mace.calculators.mace import MACECalculator

# Setup calculator
polar_calc = MACECalculator(
    model_paths="mace_mu_alpha.model",
    model_type="DipolePolarizabilityMACE",
    device="cuda",
    default_dtype="float64"
)

traj = Trajectory("test.traj", "r")
n_frames=len(traj)
alpha = np.empty((n_frames, 3, 3), dtype=float)

for i, atoms in enumerate(traj):
    atoms.calc = polar_calc
    alpha[i] = np.asarray(atoms.calc.get_property("polarizability", atoms)).reshape(3, 3)
    print(f"Frame {i} Polarizability: ", alpha[i])
traj.close()

Tip: To get the spherical polarizability (e.g. for Raman spectra), use the property "polarizability_sh" with get_property:

spherical_alpha = atoms.calc.get_property("polarizability_sh", atoms)