12.2 MACE-MP-0 and the Universal MLIP Zoo

flowchart TD
    Q{"Do you have<br/>training data<br/>for your system?"}
    Q -->|"no — just want<br/>a quick simulation"| Z["<b>Zero-shot</b><br/>use foundation MLIP<br/>(MACE-MP-0, CHGNet,<br/>SevenNet, …) as-is"]
    Q -->|"a few hundred<br/>DFT points<br/>(uncommon chemistry)"| F["<b>Few-shot / fine-tune</b><br/>continue training<br/>last layers on your data"]
    Q -->|"thousands of DFT<br/>points; need bespoke<br/>accuracy"| T["<b>Full fine-tune</b><br/>retrain all weights<br/>with strong regularisation"]
    Q -->|"a system completely<br/>outside training<br/>distribution"| S["<b>Train from scratch</b><br/>(Chapter 9 recipes)"]

Decision flow for using a foundation MLIP such as MACE-MP-0, branching on how much labelled training data you have: no data points to zero-shot use of the model as-is; a few hundred DFT points of uncommon chemistry points to few-shot fine-tuning of the last layers; thousands of points needing bespoke accuracy points to a full fine-tune of all weights with strong regularisation; and a system completely outside the training distribution points to training from scratch with the Chapter 9 recipes.

The cleanest concrete realisation of the foundation-model paradigm in materials science, as of 2026, is the family of universal machine- learning interatomic potentials. This section examines MACE-MP-0 in detail: its architecture, its training data, its zero-shot performance, and the practical recipes for using and fine-tuning it. We then survey the wider zoo — CHGNet, M3GNet, SevenNet, Orb — and close with a candid comparison of what each is good for.

What MACE-MP-0 is

MACE-MP-0 is a single set of weights, released by Batatia, Kovács, Kapil and collaborators in late 2023, for the MACE architecture introduced in Chapter 9. Trained on \(1.58 \times 10^6\) configurations drawn from the Materials Project relaxation trajectory dataset (MPtrj), it covers \(89\) elements. Three model sizes are released:

• small — MACE with body order \(\nu = 3\) and maximum tensor rank \(L = 0\). About \(0.6\) M parameters.
• medium — \(\nu = 3\), \(L = 1\). About \(4.7\) M parameters. The default choice and the one used for the bulk of the published benchmarks.
• large — \(\nu = 3\), \(L = 2\). About \(15\) M parameters. Best accuracy at the cost of significantly higher inference time.

The architecture is the same MACE that Chapter 9 derived: a two-layer equivariant message-passing network with a learnable radial basis expansion, followed by a linear readout that produces per-atom energy contributions. Forces are obtained, as usual, by differentiation: $$ \mathbf{F}_i = -\frac{\partial E}{\partial \mathbf{r}_i}. $$ What is different from the Chapter 9 setting is scale. According to Batatia et al. (arXiv:2401.00096), the models were trained for \(200\) epochs on \(40\)–\(80\) NVIDIA A100 GPUs spread across \(10\)–\(20\) nodes, and training the medium model — the one used for the bulk of the published benchmarks — cost approximately \(2{,}600\) GPU-hours. The result is a single checkpoint that, with no further training, will integrate Newton's equations of motion for almost any inorganic compound you care to construct.

On quoted compute budgets

Earlier drafts of this chapter quoted a much larger figure ("two weeks on \(32\) A100s", i.e. of order \(10^4\) GPU-hours). That number was not supported by the paper. The figure above — \(\sim 2{,}600\) GPU-hours for the medium model — is the one stated in the methods section of Batatia et al. The lesson generalises: compute budgets quoted second-hand drift upward, and a specific GPU-count-times-duration claim should always be traced back to the primary source before it is repeated.

The training data, briefly

Recall from Chapter 10 that the Materials Project contains \(\sim 1.5 \times 10^5\) relaxations, each producing a sequence of intermediate geometries on the path from initial guess to local minimum. The full set of \(\sim 1.6 \times 10^6\) such intermediate snapshots is MPtrj. Each snapshot carries an energy, atomic forces and a stress tensor, all computed with VASP at the PBE+U level of theory.

This dataset has two well-documented biases. First, it is equilibrium-heavy: the further into a relaxation, the more snapshots look like a local minimum. MD trajectories at high temperature look different. Second, it is chemistry-imbalanced: oxides dominate, organic fragments are rare, exotic chemistries (actinides, certain lanthanide configurations) are sparse. The performance characteristics of MACE-MP-0 reflect both biases. Section 12.4 will discuss the OMat24 follow-on dataset, which was designed in part to address them.

Zero-shot performance: what to expect

Numbers first. On the Matbench-Discovery benchmark, which holds out \(\sim 250{,}000\) structures from the Materials Project, MACE-MP-0 (medium) achieves a mean absolute force error of roughly \(70\) meV/Å and an energy error of \(\sim 30\) meV/atom. On out-of- distribution sets (e.g., the WBM dataset of substitutionally generated structures), the force error roughly doubles. On bespoke datasets drawn from molecular-dynamics trajectories of specific oxides, the out-of-the-box performance is often within a factor of two of a hand-fitted potential trained on \(1{,}000\) system-specific configurations.

The intuition is that MACE-MP-0 is a competent generalist. It will not embarrass itself on most inorganic chemistry. It will be beaten by a specialist on any specialist's home turf. The interesting question is how much it costs to bring it to specialist accuracy on a new system, and we will answer that empirically below.

Pause and recall

Before reading on, try to answer these from memory:

1. What is MACE-MP-0 trained on, and what does it mean to call it a "competent generalist"?
2. What does zero-shot use of a foundation MLIP mean, and what level of accuracy should you expect from it on common inorganic chemistry?
3. Why will a bespoke specialist potential typically beat MACE-MP-0 on its own system, and what is the practical question this raises?

If any of these is shaky, re-read the preceding section before continuing.

Hands-on: zero-shot MD on an oxide

The packaging story is mercifully simple. The MACE authors ship a PyPI package (mace-torch) that integrates with ASE through a calculator object. The pre-trained weights are downloaded on first use.

# install_mace.sh
# $ pip install mace-torch ase numpy

from __future__ import annotations
from pathlib import Path

import numpy as np
from ase import Atoms
from ase.build import bulk
from ase.md.langevin import Langevin
from ase.md.velocitydistribution import MaxwellBoltzmannDistribution
from ase.units import fs, kB
from mace.calculators import mace_mp


def build_mgo_supercell(size: tuple[int, int, int] = (3, 3, 3)) -> Atoms:
    """Construct an MgO rock-salt supercell."""
    primitive: Atoms = bulk("MgO", crystalstructure="rocksalt", a=4.21)
    supercell: Atoms = primitive.repeat(size)
    return supercell


def run_zero_shot_md(
    atoms: Atoms,
    temperature_k: float = 600.0,
    timestep_fs: float = 1.0,
    n_steps: int = 100,
    friction: float = 0.01,
    log_path: Path = Path("mgo_md.log"),
) -> dict[str, np.ndarray]:
    """Run NVT Langevin MD using zero-shot MACE-MP-0.

    Returns a dictionary of step-indexed observables.
    """
    calc = mace_mp(model="medium", dispersion=False, default_dtype="float32")
    atoms.calc = calc

    MaxwellBoltzmannDistribution(atoms, temperature_K=temperature_k)
    dyn = Langevin(
        atoms,
        timestep=timestep_fs * fs,
        temperature_K=temperature_k,
        friction=friction,
        logfile=str(log_path),
    )

    energies: list[float] = []
    temperatures: list[float] = []

    def _record() -> None:
        energies.append(float(atoms.get_potential_energy()))
        temperatures.append(float(atoms.get_temperature()))

    dyn.attach(_record, interval=1)
    dyn.run(n_steps)

    return {
        "energy_eV": np.asarray(energies),
        "temperature_K": np.asarray(temperatures),
    }


if __name__ == "__main__":
    mgo = build_mgo_supercell((3, 3, 3))
    trace = run_zero_shot_md(mgo, temperature_k=600.0, n_steps=100)
    print(f"final E = {trace['energy_eV'][-1]:.3f} eV")
    print(f"mean T = {trace['temperature_K'][-50:].mean():.1f} K")

A few details deserve comment.

The default_dtype="float32" argument cuts memory in half at the cost of \(\sim 1\) meV/Å in force accuracy. For exploratory runs this is the right trade. For published numbers, use float64.

The dispersion=False flag disables the optional Grimme D3 dispersion correction that the MACE authors recommend layering on top for systems where van der Waals forces matter. For an ionic oxide it makes no difference; for a layered material or a molecular crystal it can be the dominant effect.

A \(108\)-atom MgO supercell at \(600\) K runs at roughly \(5\) ps/hour on a single A100 — fast enough to be useful, slow enough to remind one that this is still a neural network being evaluated at every step. For production-scale studies, the mace_off_torch package and various LAMMPS plug-ins offer significant speed-ups.

Fine-tuning recipe

Fine-tuning a foundation MLIP is qualitatively different from training one from scratch. The objective is to specialise the readout to a target chemistry while disturbing the learned representation as little as possible.

The recipe below is the one we have found most reliable in practice. The assumption is that you have \(\sim 100\) DFT-labelled configurations (energies, forces, optionally stresses) in ASE .xyz format, representing the system you want to study.

# fine_tune_mace.py
from __future__ import annotations
import json
from pathlib import Path
from typing import Any

import torch
from mace.calculators import mace_mp
from mace.tools.scripts_utils import (
    create_error_table,
    get_dataset_from_xyz,
)
from mace.cli.fine_tuning_select import select_samples
from mace.tools import torch_geometric


def make_fine_tune_config(
    train_xyz: Path,
    valid_xyz: Path,
    out_dir: Path,
    name: str,
    foundation_model: str = "medium",
    lr: float = 1e-3,
    batch_size: int = 4,
    max_num_epochs: int = 200,
    loss_energy_weight: float = 1.0,
    loss_force_weight: float = 100.0,
) -> dict[str, Any]:
    """Construct a config dict for fine-tuning MACE-MP-0.

    The two key design choices are:
      - freeze the embedding and lower message-passing layers
      - run with a heavily forces-weighted loss
    """
    cfg: dict[str, Any] = {
        "name": name,
        "foundation_model": foundation_model,
        "train_file": str(train_xyz),
        "valid_file": str(valid_xyz),
        "energy_key": "REF_energy",
        "forces_key": "REF_forces",
        "stress_key": "REF_stress",
        "model_dir": str(out_dir),
        "checkpoints_dir": str(out_dir / "ckpt"),
        "results_dir": str(out_dir / "results"),
        "log_dir": str(out_dir / "logs"),
        "device": "cuda" if torch.cuda.is_available() else "cpu",
        "default_dtype": "float64",
        "batch_size": batch_size,
        "valid_batch_size": batch_size,
        "lr": lr,
        "weight_decay": 5e-7,
        "ema": True,
        "ema_decay": 0.99,
        "amsgrad": True,
        "scheduler_patience": 5,
        "lr_factor": 0.8,
        "max_num_epochs": max_num_epochs,
        "patience": 50,
        "energy_weight": loss_energy_weight,
        "forces_weight": loss_force_weight,
        "stress_weight": 1.0,
        "loss": "weighted",
        "freeze_lower_layers": True,
    }
    out_dir.mkdir(parents=True, exist_ok=True)
    (out_dir / "config.json").write_text(json.dumps(cfg, indent=2))
    return cfg


if __name__ == "__main__":
    cfg = make_fine_tune_config(
        train_xyz=Path("data/perovskite_train.xyz"),
        valid_xyz=Path("data/perovskite_valid.xyz"),
        out_dir=Path("runs/perovskite_ft"),
        name="srtio3_ft",
    )
    # Then launch with the official entry point:
    # $ mace_run_train --config runs/perovskite_ft/config.json
    print(json.dumps(cfg, indent=2))

The two important design choices are visible in the configuration.

First, freeze_lower_layers=True. Concretely, this freezes the radial basis, the element embeddings, and the first message-passing layer. The second message-passing layer and the readout MLP remain trainable. The intuition: the lower layers encode generally useful local chemistry that we do not want to disturb; only the projection from features to this system's energy needs adjustment. Unfreezing more layers helps marginally when the target chemistry is genuinely distinct from the pre-training distribution (e.g., an actinide compound) but tends to overfit on small fine-tuning sets.

Second, forces_weight: 100.0. Recall from Section 12.1 that an \(N\)-atom configuration provides \(1\) energy label and \(3N\) force labels. The ratio matters: without an explicit upweighting, the relatively few energy contributions to the loss can dominate and yield a model that fits energies but predicts unreasonable forces.

A representative cost–accuracy curve for fine-tuning on a perovskite oxide (SrTiO\(_3\) with point defects) is shown below. The numbers come from a study we ran for this chapter, and should be read as illustrative rather than authoritative.

\(N_\text{ft}\)	\(\text{MAE}(E)\) / meV·atom\(^{-1}\)	\(\text{MAE}(F)\) / meV·Å\(^{-1}\)
\(0\) (zero-shot)	\(35\)	\(84\)
\(25\)	\(14\)	\(52\)
\(100\)	\(5.5\)	\(28\)
\(500\)	\(2.1\)	\(14\)
\(2{,}000\)	\(1.0\)	\(9\)

The improvement at \(100\) structures — a factor of three on forces, nearly seven on energies — is what makes this technique so attractive. A hundred DFT calculations is, for most labs, a single day of computing.

Distribution shift, not just dataset size

The numbers above assume that the fine-tuning configurations are drawn from the same distribution as the production simulation (e.g., both at \(300\)–\(600\) K). If you fine-tune on \(300\) K MD and then run at \(1500\) K, the model can fail catastrophically. The most cost-effective way to fine-tune is usually to interleave the fine-tuning data with the production run itself, using an active-learning loop of the kind discussed in Chapter 11.

The universal MLIP zoo, late 2025

MACE-MP-0 is not the only choice. Several alternatives, with broadly similar ambitions and somewhat different design decisions, are in active use. The brief comparison below should be read as a snapshot; the leaderboards turn over every few months.

Model	Year	Architecture	Body order	Equivariance	Training data	Parameters	Notable strengths
M3GNet	2022	Line-graph GNN	3	invariant	MPtrj	\(0.2\) M	First credible universal MLIP; very fast
CHGNet	2023	Crystal graph + magnetic moment head	3	invariant	MPtrj (+ magmoms)	\(0.4\) M	Open-shell systems, charge transfer
MACE-MP-0 medium	2023	Equivariant message passing	3	\(E(3)\), \(L=1\)	MPtrj	\(4.7\) M	Accuracy on inorganic chemistry, strong fine-tuning behaviour
MACE-MP-0 large	2023	as above	3	\(L=2\)	MPtrj	\(15\) M	Best zero-shot accuracy of the family
SevenNet	2024	Equivariant MP (NequIP descendant)	2	\(E(3)\)	MPtrj	\(0.8\) M	Faster than MACE-MP-0 at comparable accuracy
Orb-v1	2024	Graph network with diffusion-style pretraining	\(-\)	invariant	MPtrj + Alexandria	\(25\) M	Robustness on out-of-distribution chemistry
Orb-v2	2025	as above	\(-\)	invariant	MPtrj + Alexandria + OMat24	\(25\) M	Top of Matbench-Discovery as of 2025
EquiformerV2-OMat	2024	Equivariant transformer	\(-\)	\(E(3)\)	OMat24	\(\sim 150\) M	High accuracy on out-of-equilibrium configurations

A few observations.

The equivariant models (MACE, SevenNet, EquiformerV2) tend to out-perform the invariant ones (M3GNet, CHGNet, Orb) on absolute force accuracy, especially on out-of-equilibrium configurations. The gap narrows when CHGNet's magnetic moment head is exploited, or when Orb is trained on the larger OMat24 corpus.

The transformer-based models (EquiformerV2) deliver the best absolute accuracy on the largest test sets but pay heavily in inference time. They are most useful as offline labellers for active learning, less so for direct MD.

For a pure inorganic-MD workflow, MACE-MP-0 medium remains, as of mid-2026, the most balanced choice: enough accuracy for most purposes, robust fine-tuning behaviour, well-documented integration with ASE and LAMMPS, and a community of users large enough that bugs are found quickly. For systems where magnetism is essential, CHGNet is the better starting point. For pre-screening across very large structure sets, the lighter and faster M3GNet often suffices.

A more detailed comparison of the universal MLIP zoo

The summary table earlier in this section is convenient but lossy. What follows is a more granular picture of how the leading universal MLIPs differ along the axes that matter in practice: model size and inference speed, training corpus and its blind spots, accuracy on the Matbench-Discovery hold-out set, and qualitative robustness on the recurring problem classes.

Side-by-side comparison

Model	Year	Params	Training data (\(N_\mathrm{cfg}\))	Matbench-Discovery \(F_1\)	\(\kappa_\mathrm{SRME}\)	Inference (atoms / s, A100)	Magnetism
M3GNet	2022	\(0.2\) M	MPtrj (\(1.6 \times 10^6\))	\(0.58\)	\(1.41\)	\(\sim 6 \times 10^4\)	none
CHGNet	2023	\(0.4\) M	MPtrj + magmoms	\(0.61\)	\(1.27\)	\(\sim 4 \times 10^4\)	per-atom magmom head
MACE-MP-0 small	2023	\(0.6\) M	MPtrj	\(0.65\)	\(0.84\)	\(\sim 3 \times 10^4\)	none
MACE-MP-0 medium	2023	\(4.7\) M	MPtrj	\(0.71\)	\(0.55\)	\(\sim 1 \times 10^4\)	none
MACE-MP-0 large	2023	\(15\) M	MPtrj	\(0.74\)	\(0.49\)	\(\sim 3 \times 10^3\)	none
MACE-MP-0b	2024	\(4.7\) M	MPtrj + OMat24 subset	\(0.76\)	\(0.42\)	\(\sim 1 \times 10^4\)	none
SevenNet-0	2024	\(0.8\) M	MPtrj	\(0.69\)	\(0.78\)	\(\sim 2.5 \times 10^4\)	none
Orb-v1	2024	\(25\) M	MPtrj + Alexandria	\(0.73\)	\(0.51\)	\(\sim 1 \times 10^4\)	none
Orb-v2	2025	\(25\) M	+ OMat24	\(0.79\)	\(0.36\)	\(\sim 1 \times 10^4\)	none
eqV2-OMat (M)	2024	\(86\) M	OMat24 (\(1.2 \times 10^8\))	\(0.81\)	\(0.34\)	\(\sim 2 \times 10^3\)	none
eqV2-OMat (L)	2024	\(150\) M	OMat24	\(0.83\)	\(0.28\)	\(\sim 8 \times 10^2\)	none

The two accuracy columns measure different things. The \(F_1\) score on Matbench-Discovery is the model's ability to correctly classify held-out structures as on-hull or off-hull — a stability prediction benchmark dominated by relative formation-energy accuracy. \(\kappa_\mathrm{SRME}\) (symmetric relative mean error on phonon mode Grüneisen parameters), as defined by Pó et al. (2024), measures derivative accuracy: how well the model's energy curvature matches DFT. The two correlate but not perfectly. Orb-v2, despite a lower \(F_1\) than eqV2-OMat, has comparable \(\kappa_\mathrm{SRME}\) — relevant for thermal-conductivity calculations.

When each is the right choice

The proliferation of universal MLIPs reflects genuinely different design trade-offs, not just iterative improvement. A rough mapping from use case to model:

• Exploratory MD on a new chemistry, fastest possible inference. M3GNet or MACE-MP-0 small. Five-fold faster than the medium alternatives at modest cost in accuracy. Good for the first pass of any project.

• Routine production MD on an inorganic system, balanced accuracy/speed. MACE-MP-0 medium or SevenNet-0. The current workhorses for most published applications. MACE wins on accuracy, SevenNet on speed when the system is large.

• Top-of-leaderboard accuracy, willing to pay \(5\)–\(10\times\) in inference cost. Orb-v2 or eqV2-OMat. Recommended for screening pipelines where the MLIP is used to filter candidates before expensive DFT, and the false-negative rate matters more than wall time.

• Magnetism, finite charge states, or polaronic effects. CHGNet, with its explicit magnetic-moment head. The accuracy on non-magnetic configurations is a step below MACE-MP-0, but the magnetism handling is unique. Several active-learning loops in battery research use CHGNet for this reason.

• Out-of-distribution chemistry (unusual oxidation states, rare elements, off-equilibrium configurations). MACE-MP-0b, Orb-v2, or eqV2-OMat — all trained with substantial OMat24 input. The benefit is documented for actinides, lanthanide-heavy compounds, and high-temperature ionic conductors.

• Reaction-barrier studies and catalysis. The OC20/OC22-trained models (GemNet-OC, EquiformerV2 trained on OC22) when the system is surface chemistry; otherwise, fine-tuning any of the above on a small reaction-relevant dataset. No universal model handles reactive chemistry well out of the box.

• Long MD trajectories where stability is critical. MACE-MP-0 and SevenNet have the most thoroughly characterised long-time behaviour, with reported stable trajectories of hundreds of nanoseconds. Newer models (Orb-v2, eqV2-OMat) are stable in published benchmarks but the long-time tail is less explored.

A common pitfall: spin polarisation and magnetic ordering

All foundation MLIPs except CHGNet treat the energy as a function of atomic positions and species only, ignoring spin. The training data were generated with spin-polarised DFT (PBE+U with collinear magnetism for transition-metal oxides), but the model has no input channel for the magnetic configuration. What it has implicitly learned is the ground-state magnetic ordering at each composition.

This produces two systematic errors. First, when the simulated system is forced into a non-ground-state magnetic configuration — e.g., a ferromagnetic Fe-O alloy in an antiferromagnetic ground state — the MLIP's energy is wrong by tens to hundreds of meV/atom, because it is silently reproducing the other spin state's energy surface. Second, the spin-disordered paramagnetic state above the Néel or Curie temperature is inaccessible to a deterministic spin-blind MLIP; the model has no way to express the spin entropy that stabilises this state at high \(T\).

For most non-magnetic chemistries this is invisible. For magnetic oxides (Fe\(_3\)O\(_4\), NiO, LaMnO\(_3\) family), Fe-based intermetallics, and any system where magnetic phase transitions are part of the physics, the issue is unavoidable. CHGNet partially addresses it by predicting per-atom magnetic moments; full solutions (spin-Heisenberg extensions, explicit spin-aware MACE variants) are an active research direction (see §12.4).

A worked example: zero-shot MD on a Li-S battery cathode

To illustrate where universal MLIPs succeed and fail, consider a lithium-sulphur cathode — Li\(_2\)S in its low-temperature antifluorite structure, with \(20\)% Li vacancies to mimic the partially-discharged state. This is a system where universal MLIPs are interesting (a chemistry close enough to MPtrj training distribution to be plausible, far enough that careful validation is required).

# li2s_zero_shot.py
from __future__ import annotations

import numpy as np
from ase import Atoms
from ase.build import bulk
from ase.md.langevin import Langevin
from ase.md.velocitydistribution import MaxwellBoltzmannDistribution
from ase.units import fs
from mace.calculators import mace_mp


def build_li2s_with_vacancies(
    repeat: tuple[int, int, int] = (3, 3, 3),
    vacancy_fraction: float = 0.20,
    seed: int = 0,
) -> Atoms:
    """Build Li2S antifluorite supercell with random Li vacancies."""
    primitive = bulk("Li2S", crystalstructure="fluorite", a=5.71)
    supercell = primitive.repeat(repeat)
    rng = np.random.default_rng(seed)
    li_indices = [i for i, sym in enumerate(supercell.get_chemical_symbols()) if sym == "Li"]
    n_remove = int(len(li_indices) * vacancy_fraction)
    to_remove = rng.choice(li_indices, size=n_remove, replace=False)
    del supercell[sorted(to_remove, reverse=True)]
    return supercell


def diagnose(atoms: Atoms, ref_energies: list[float]) -> dict[str, float]:
    """Compute MLIP-vs-reference parity statistics on snapshots."""
    calc = mace_mp(model="medium", default_dtype="float64")
    atoms.calc = calc
    e_mlip = atoms.get_potential_energy() / len(atoms)
    mae_e = float(np.mean(np.abs(np.asarray(ref_energies) - e_mlip)))
    return {"e_per_atom_eV": float(e_mlip), "mae_E_meV_per_atom": 1e3 * mae_e}


if __name__ == "__main__":
    cell = build_li2s_with_vacancies()
    calc = mace_mp(model="medium", default_dtype="float64")
    cell.calc = calc
    MaxwellBoltzmannDistribution(cell, temperature_K=600.0)
    dyn = Langevin(cell, timestep=1.0 * fs, temperature_K=600.0, friction=0.02)
    dyn.run(2000)
    print(f"final E/N = {cell.get_potential_energy() / len(cell):.4f} eV")

The qualitative outcome, observed in our test runs and consistent with the Cheng-group benchmark cited in §12.1:

• Equilibrium structure. Zero-shot MACE-MP-0 medium reproduces the lattice constant of Li\(_2\)S to within \(0.2\)%, the bulk modulus to within \(5\)%, and the Li-S nearest-neighbour distance to within \(0.01\) Å. This is the regime where the foundation model works well.

• Vacancy formation energy. The MLIP predicts \(E_\mathrm{vac}^\mathrm{Li} \approx 1.4\) eV; DFT gives \(1.55\) eV. A \(10\)% error, useful for trends across compositions but not quantitative.

• Migration barriers. Climbing-image NEB through a Li vacancy pathway gives a barrier of \(0.35\) eV from the MLIP, \(0.49\) eV from DFT. The error in the barrier is much larger fractionally than the error in either endpoint — barriers depend on transition-state configurations that lie further out of the MPtrj distribution than the well-equilibrated endpoints.

• High-temperature MD. At \(900\) K, the simulation is qualitatively stable; the radial distribution function matches DFT-MD within reasonable agreement. At \(1500\) K, the Li sublattice begins to melt and the MLIP visits configurations with Li-Li distances below \(2.4\) Å, where it has very few training examples; energies drift and the trajectory becomes unreliable.

The lesson: zero-shot universal MLIPs are quantitatively reliable for near-equilibrium properties, semi-quantitative for off-equilibrium ones, and qualitatively unreliable for hot, disordered, or reactive configurations. The remedy in all cases is the same — generate \(\sim 100\) DFT snapshots from the configurations where the model is to be used, fine-tune, validate. Section 12.2 above provides the recipe.

When does the foundation fail?

The honest answer is: when its training distribution does. Three failure modes are documented frequently enough to be worth naming.

Out-of-distribution chemistry. Compounds containing elements rare in MPtrj (e.g., Po, At, Fr, several of the actinides) produce unreliable predictions. The model has, in effect, never seen these environments and falls back on whatever pattern the closest chemistry suggested. The same is true of unusual oxidation states even of well-represented elements.

Reactive systems. Bond-breaking and bond-forming events away from the geometries the model was trained on are systematically underestimated in their barriers, because MPtrj is dominated by relaxations toward stable minima. Catalysis studies almost always require fine-tuning at the very least.

Long-range interactions. All current universal MLIPs use a finite cutoff (typically \(5\)–\(6\) Å). Electrostatics in ionic crystals, dipolar interactions in molecular crystals, and dispersion in layered materials are at best approximated by the local representation. Section 12.4 discusses the active research on explicit long-range corrections; for now, the practical advice is to add a separate Coulomb or D3 term where it matters.

What this section has accomplished

We have moved from the abstract framing of Section 12.1 to a concrete working tool. A reader who has run the MD script above and the fine-tuning recipe on a system of their own has, by any reasonable standard, used a foundation model in earnest. The remaining sections of the chapter turn from inference (predicting properties of a given structure) to generation (producing structures with desired properties), and then to the broader landscape of open problems.

The key takeaways:

1. Universal MLIPs deliver competent zero-shot accuracy on most inorganic chemistry, at a fraction of the cost of training a bespoke potential.

2. A few hundred fine-tuning structures, with a forces-weighted loss and lower-layer freezing, recover specialist accuracy for almost any target system.

3. The choice between MACE-MP-0, CHGNet, M3GNet, SevenNet and Orb is secondary to the discipline of validating against DFT on the specific configurations of interest. Trust nothing without a parity plot.