ML for Materials · Research Series

Why symmetry-aware machine learning potentials learn better physics — from core concepts to implementation.

Series

Physical laws governing molecules are symmetric: rotating a crystal doesn’t change its energy; reflecting a lattice shouldn’t surprise your model. Encoding that symmetry directly into the network architecture — rather than hoping the network learns it from millions of training examples — is the foundational idea behind modern ML interatomic potentials and the reason equivariant networks beat invariant ones on every benchmark that matters for materials.

Why Symmetry Matters

The physical laws governing molecules are symmetric. Encoding that symmetry directly into a neural network is the foundational idea behind modern ML potentials.

C H H H
⟶ R ⟶
C H H H
E(R·{ri}) = E({ri}) — Energy is invariant F({ri}) → RF({R−1ri}) — Forces are equivariant
🔄

Rotational Invariance

When you rotate a molecule, its total energy does not change. A physically correct model must respect this — not learn it from millions of rotated examples. E(R·{ri}) = E({ri})

⃗F

Force Equivariance

Forces are vectors. When you rotate the molecule, the force vectors rotate with it. The model output must transform consistently with the input transformation. F → RF(R-1·{ri})

📉

Data Efficiency

An equivariant network is 10–100× more data-efficient than one that must learn symmetry from examples. Critical when DFT data is expensive to generate.

Key Insight

A standard neural network wastes model capacity learning that “a water molecule pointing left has the same energy as one pointing right.” Equivariant networks encode this as a mathematical constraint, freeing every parameter to learn actual chemistry.

How Equivariance is Built In

  1. Represent atoms as tensors, not scalars. Each atom carries not just a scalar feature (charge, mass) but vector and higher-order tensor features that transform correctly under rotation — using irreducible representations (irreps) of SO(3).
  2. Encode geometry with spherical harmonics. Pairwise directions between atoms are projected onto spherical harmonics Ylm(r̂ij), the natural rotationally-equivariant basis functions on the sphere.
  3. Combine via tensor products. Features are combined using Clebsch–Gordan tensor products that preserve equivariance at every layer. SO(3) symmetry is maintained throughout the entire message passing process.
  4. Extract invariant energy output. By contracting all tensor indices at the output layer, we obtain a scalar (invariant) total energy. Forces are then the analytical gradient — automatically equivariant.

Architecture Timeline

Each generation of ML potential architectures has encoded more physical symmetry, dramatically improving accuracy and data efficiency.

ModelYearSymmetryKey MechanismData Efficiency
SchNet2017 Invariant Continuous filter convolutions on scalar distances only Baseline — needs large datasets
DimeNet2020 Invariant Adds bond angles via directional message passing Better than SchNet, limited angular expressivity
NequIP2022 Equivariant E(3) symmetry via tensor products of irreps; spherical harmonics edge features ~10× more efficient than SchNet; near-DFT accuracy
MACE2022 Equivariant+ Higher-order tensor products capturing 3- and 4-body interactions explicitly State-of-the-art; fast inference; universal models (89 elements)
MACE-MP-0 — Universal Potential

MACE-MP-0 is trained on the Materials Project database and achieves competitive accuracy across all 89 elements of the periodic table with a single model. It can be used as a zero-shot calculator for any inorganic material, or fine-tuned on domain-specific DFT data with very few additional calculations.

Invariant vs. Equivariant — The Key Difference

📐

Invariant Networks (SchNet / DeepMD style)

Operate only on scalar features (distances). Directional information is discarded after computing descriptors. Angular information is limited and higher-order interactions are expensive. Must learn directional physics implicitly from data — information loss at every layer.

🧮

Equivariant Networks (NequIP / MACE style)

Vector and tensor features propagate through the network intact. SO(3) symmetry encoded via spherical harmonics and tensor products. Directional information preserved at every message passing step — full geometric information preserved.

The two key mathematical statements E(R·{ri}) = E({ri})     ← Energy invariant under rotation R
F({ri}) → RF({R−1·ri})   ← Forces equivariant (rotate with molecule)

h(l)i = Σj Wl · (h(l-1)j ⊗ Yl(r̂ij))    ← MACE tensor product message passing

Code Examples

Practical Python snippets for using equivariant ML potentials in your research — from installation to running MD simulations.

1. Install & Load MACE-MP-0 (Universal Potential)

# Install MACE
pip install mace-torch

from mace.calculators import mace_mp
from ase.build import bulk
import numpy as np

# Load the universal MACE-MP-0 potential (downloads ~50 MB model)
calc = mace_mp(model="medium", dispersion=False, default_dtype="float32")

# Create a silicon crystal structure using ASE
si = bulk("Si", "diamond", a=5.43)
si.calc = calc

# Compute energy and forces — no DFT required!
energy = si.get_potential_energy()   # eV
forces = si.get_forces()             # eV/Å, shape (N_atoms, 3)
stress = si.get_stress()             # eV/ų

print(f"Energy: {energy:.4f} eV")
print(f"Forces shape: {forces.shape}")
print(f"Max force: {np.max(np.abs(forces)):.4f} eV/Å")

2. Structure Relaxation with MACE

from ase.optimize import BFGS
from ase.io import read, write

# Load your structure (CIF, POSCAR, xyz, etc.)
atoms = read("my_structure.cif")
atoms.calc = calc

# Relax atomic positions and cell
opt = BFGS(atoms, logfile="relax.log")
opt.run(fmax=0.01)   # Converge forces to 0.01 eV/Å

# Save relaxed structure
write("relaxed.cif", atoms)
print(f"Relaxed energy: {atoms.get_potential_energy():.4f} eV")

3. Molecular Dynamics with MACE

from ase.md.langevin import Langevin
from ase.md.velocitydistribution import MaxwellBoltzmannDistribution
from ase import units

# Set initial velocities at 300 K
MaxwellBoltzmannDistribution(atoms, temperature_K=300)

# Langevin NVT thermostat
dyn = Langevin(
    atoms,
    timestep=1.0 * units.fs,    # 1 fs timestep
    temperature_K=300,
    friction=0.01 / units.fs,
)

# Run 1000 steps (1 ps) of MD
print("Step | Energy (eV) | Temp (K)")
for i in range(1000):
    dyn.run(1)
    if i % 100 == 0:
        E = atoms.get_potential_energy()
        T = atoms.get_temperature()
        print(f"{i:5d} | {E:12.4f} | {T:8.1f}")

4. Fine-tune NequIP on Your Own DFT Data

# nequip_train.yaml — minimal config for fine-tuning
root: results/my_system
run_name: nequip_finetune

dataset: ase
dataset_file_name: my_dft_dataset.xyz   # Extended XYZ with E+F+stress
n_train: 800
n_val:   100
batch_size: 5

# Model architecture
num_layers: 4
l_max: 2              # Max angular momentum (use 2 or 3)
num_features: 32      # Channels per irrep
r_max: 5.0            # Cutoff radius in Å
parity: true

# Training
max_epochs: 1000
learning_rate: 0.005
loss_coeffs:
  total_energy: 1
  forces: 10           # Weight forces more heavily
Tip for Educators

The ASE (Atomic Simulation Environment) library provides a unified interface. Students can swap between DFT calculators (VASP, Quantum ESPRESSO) and MACE with a single line change — making it excellent for comparing ML vs. DFT results directly.

Self-Assessment Quiz

Test your understanding of equivariant neural networks. Select an answer to see instant feedback.

1. When a molecule is rotated in space, what happens to its total potential energy?
2. What is the main advantage of equivariant networks over invariant networks like SchNet?
3. Which mathematical objects are used in equivariant networks to represent directions between atoms?
4. What makes MACE superior to NequIP for capturing many-body interactions?
5. MACE-MP-0 is described as a “universal potential.” What does this mean?
0/5

Glossary

Key terms for understanding equivariant neural networks in computational materials science.

Equivariance
A property of a function where applying a transformation to the input results in a predictable transformation of the output. For neural networks in chemistry: rotating atomic positions leads to rotated force vectors. Formally: f(T·x) = T·f(x) for transformation T.
Invariance
A property where applying a transformation to the input leaves the output unchanged. Total energy is invariant under rotation, translation, and permutation of identical atoms. Formally: f(T·x) = f(x). Invariance is a special case of equivariance where the output is a scalar.
SO(3)
The special orthogonal group in 3 dimensions — the mathematical group of all rotations in 3D space. Equivariant networks are designed so their internal representations transform according to the irreducible representations (irreps) of SO(3). Including reflections gives O(3).
Spherical Harmonics Ylm
Orthogonal basis functions defined on the unit sphere, indexed by angular momentum l and magnetic quantum number m. They are the natural equivariant basis for encoding 3D directional information. l=0 gives scalars, l=1 gives vectors, l=2 gives rank-2 tensors. MACE uses l up to 3 or 4.
Irreducible Representations (Irreps)
The fundamental building blocks of how a group acts on vector spaces. For SO(3), the irreps are labeled by angular momentum l. Features in equivariant networks are decomposed into irreps so that rotations act on them in a well-defined, mathematically consistent way.
Clebsch–Gordan Tensor Product
The equivariant operation used to combine two sets of irrep features into new features. It generalizes matrix multiplication to the full tensor structure required to maintain SO(3) equivariance. This is the key computational operation in NequIP and MACE.
MACE (Multi-Atomic Cluster Expansion)
State-of-the-art equivariant ML potential architecture (2022, Batatia et al.). Extends NequIP by computing higher-order equivariant tensor products within each message passing step, enabling explicit 3-body and 4-body correlations. MACE-MP-0 is a universal foundation model for materials.
Machine Learning Interatomic Potential (MLIP / MLP)
A neural network trained to predict the potential energy surface of a molecular or crystalline system, enabling molecular dynamics and structure optimization at a fraction of DFT cost. Modern MLIPs like MACE achieve near-DFT accuracy with orders-of-magnitude speedup.
DFT (Density Functional Theory)
The gold-standard quantum mechanical method for computing electronic structure, energies, and forces. DFT is accurate but computationally expensive (scales ~O(N³)). MLIPs trained on DFT data aim to reproduce DFT accuracy at classical force field cost (O(N)).
Message Passing Neural Network (MPNN)
A graph neural network architecture where atoms are nodes and bonds/interactions are edges. At each layer, atoms “pass messages” to their neighbors, aggregating information. Equivariant MPNNs (NequIP, MACE) pass full tensor messages rather than scalar messages, preserving geometric information.
Materials Project
A large open-access database of DFT-computed materials properties for over 150,000 inorganic compounds. MACE-MP-0 is trained on this dataset, which is why it generalizes across most elements of the periodic table.
Equivariant Networks MACE NequIP SO(3) Symmetry ML Potentials Spherical Harmonics Materials Project