MACE-Osaka24: Universal Interatomic Potential

• MACE-Osaka24 is a foundational model that unifies molecular and crystalline simulations by merging heterogeneous quantum datasets through Total Energy Alignment (TEA).
• Its architecture leverages an E(3)-equivariant message-passing neural network, ensuring robust predictions of energies, forces, and stresses.
• Benchmark studies demonstrate state-of-the-art performance in organic reactions, biaryl torsions, and inorganic lattice constants, achieving chemical accuracy.

Searching arXiv for the MACE-Osaka24 paper and closely related MACE architecture/context papers. MACE-Osaka24 is a universal, open-source machine-learning interatomic potential built on the MACE architecture and trained on a single unified dataset spanning both inorganic crystalline systems and organic molecular systems. Its defining technical contribution is the use of Total Energy Alignment (TEA) to merge heterogeneous quantum-chemical datasets computed with different electronic-structure methods into one coherent training set, thereby enabling a single neural potential to model molecular and crystalline chemistry within a common potential-energy framework (Shiota et al., 2024).

1. Definition and scope

MACE-Osaka24 denotes the first open-source neural network potential based on a unified dataset covering both molecular and crystalline systems, using the MACE architecture developed by Batatia et al. The model is positioned as a foundation model for atomistic simulations: a general-purpose potential intended for reuse across chemistry and materials tasks without requiring domain-specific retraining for each application. Two released variants are described, “MACE-Osaka24-small” and “MACE-Osaka24-large,” following the naming convention of earlier MACE foundation models (Shiota et al., 2024).

The model emerges from the convergence of two earlier domain-specialized MACE lines. MACE-MP-0 was trained for inorganic materials on MPtrj, while MACE-OFF23 was trained for organic chemistry on OFF23. MACE-Osaka24 differs in that it uses one single MACE network to describe both domains and is trained on a merged inorganic–organic dataset rather than on separate domain-specific corpora (Shiota et al., 2024).

Within the broader evolution of interatomic potentials, MACE itself belongs to the progression from Gaussian Approximation Potentials (GAP) to the Atomic Cluster Expansion (ACE) and then to MACE as a multilayer neural-network extension of ACE. In that lineage, MACE retains the locality, symmetry structure, and systematic many-body character of ACE while replacing the linear site-energy model with an E(3)-equivariant message-passing architecture (Bernstein, 2024).

2. Total Energy Alignment

The central methodological obstacle addressed by MACE-Osaka24 is that its two principal source datasets inhabit different total-energy conventions. MPtrj uses PBE with VASP, a plane-wave basis, and PAW pseudopotentials, whereas OFF23 uses $\omega$B97M-D3(BJ) with def2-TZVPPD in Psi4 and an all-electron treatment. Direct concatenation would therefore mix incompatible potential-energy surfaces, since the resulting total energies differ because of functional choice, basis representation, and core-electron treatment (Shiota et al., 2024).

TEA resolves this by mapping one dataset onto the energy scale of the other with almost no redundant recalculation. The procedure has two components. The first is Inner Core Energy Alignment (ICEA), which compensates composition-dependent constant shifts arising from different isolated-atom reference energies. The second is Atomization Energy Correction (AEC), which applies a global linear scaling to the atomization energy in order to correct remaining systematic discrepancies between methods (Shiota et al., 2024).

For a system of $N$ atoms with element types $P_i$, the atomization energy is defined as

$$E_{\mathrm{at}} = \sum_{i=1}^N E_i^{P_i} - E_{\mathrm{tot}},$$

where $E_i^{P_i}$ is the isolated-atom energy of species $P_i$ and $E_{\mathrm{tot}}$ is the total energy. Under ICEA, one assumes approximate equality of atomization energies between two methods and obtains the aligned total energy through a composition-dependent offset,

$$E_{\mathrm{tot}}^{[1]} = E_{\mathrm{tot}}^{[2]} + \sum_{i=1}^N\left(E_i^{P_i,[1]} - E_i^{P_i,[2]}\right).$$

AEC then introduces a global scalar $a$ such that

$E_{\mathrm{at}}^{[1]} = a\,E_{\mathrm{at}}^{[2]},$

which yields

$N$0

Because forces are total-energy gradients, the same scaling applies to forces,

$N$1

This makes TEA a global linear mapping in both energies and forces: a composition-dependent shift plus a global scale factor (Shiota et al., 2024).

The scalar $N$2 is calibrated on QM9-based cross-method comparisons, specifically QM9VASP, QM9Psi4, and QM9ADF. The reported validation shows that for QM9VASP versus QM9ADF, the raw atomization-energy RMSE of 0.3258 eV is reduced to a total-energy RMSE of 0.0992 eV after ICEA + AEC. For QM9VASP versus QM9Psi4, ICEA alone still leaves an RMSE of about 4.20 eV, while ICEA + AEC reduces the total-energy RMSE to about 0.84 eV. This establishes TEA as a practical mechanism for combining heterogeneous datasets without recomputing all structures at a common level of theory (Shiota et al., 2024).

3. Unified dataset and model construction

The unified training corpus is formed from MPtrj and OFF23. MPtrj supplies the inorganic materials domain and was originally computed with PBE in VASP. OFF23 supplies the organic molecular domain and itself combines SPICE, QMugs, water clusters, and tripeptides, computed with $N$3B97M-D3(BJ)/def2-TZVPPD in Psi4. Before alignment, the D3(BJ) contribution is removed from OFF23 to avoid double counting when dispersion corrections are later added in simulation workflows (Shiota et al., 2024).

After TEA, the aligned OFF23 data are merged with MPtrj into the TEA-MPtrj/OFF23 dataset. The resulting corpus is intended to support one training task rather than a multi-head or multi-domain objective with separate energy scales. A plausible implication is that TEA converts dataset heterogeneity from an optimization problem into a representation problem: once the label spaces are aligned, the remaining question is whether a single equivariant architecture has sufficient capacity to represent both chemical domains (Shiota et al., 2024).

MACE-Osaka24 is trained with the MACE architecture implemented in mace v0.3.6. The model uses a cutoff radius of 4.5 Å for constructing the local atomic graph. As in standard MACE, the total energy is written as a sum of per-atom contributions,

$N$4

and forces are obtained from the analytic gradient,

$N$5

The loss is the standard MACE multi-term objective over energies, forces, and virial stresses, schematically

$N$6

Spin-polarized isolated-atom energies from VASP PBE are used as the atomic reference energies for all species, including those appearing in OFF23, so that atomization-energy definitions remain consistent across the merged dataset (Shiota et al., 2024).

Training is performed on 32 A100 GPUs in parallel. The paper reports that the energy, force, and stress RMSE all decrease smoothly over approximately 200 epochs, with the large model consistently achieving lower error than the small model. This training behavior suggests that the merged dataset does not introduce obvious optimization pathologies after TEA alignment (Shiota et al., 2024).

4. Benchmark performance across domains

The biaryl torsion benchmark provides the clearest molecular comparison against specialized models. On 78 drug-like biaryl rotors with CCSD(T1)*/CBS barrier references, MACE-Osaka24-large attains a barrier-height MAE of 0.457 kcal/mol, while MACE-Osaka24-small attains 0.695 kcal/mol. These results remain within chemical-accuracy territory and substantially outperform MACE-MP-0-large at 1.909 kcal/mol and MACE-MP-0-small at 3.386 kcal/mol. They are only slightly less accurate than the domain-specialized MACE-OFF23-large at 0.403 kcal/mol, despite MACE-Osaka24 being trained simultaneously on a broad inorganic domain (Shiota et al., 2024).

On Transition1x, which evaluates reaction energies and activation barriers on 10,073 organic reactions using fixed initial-state, transition-state, and final-state geometries, MACE-Osaka24 becomes the strongest reported model among those compared. MACE-Osaka24-large reaches a reaction-energy MAE of 0.404 eV and a barrier MAE of 0.265 eV, while MACE-Osaka24-small reaches 0.457 eV and 0.336 eV. Both outperform MACE-MP-0 and MACE-OFF23 of corresponding size. This suggests that cross-domain pretraining may improve transfer into off-equilibrium molecular configurations rather than merely preserving equilibrium accuracy (Shiota et al., 2024).

For inorganic crystals, the lattice-constant benchmark shows that MACE-Osaka24 preserves near-state-of-the-art structural accuracy. The reported mean absolute error in equilibrium lattice constants is 0.020 Å for MACE-Osaka24-small and 0.018 Å for MACE-Osaka24-large, compared with 0.012 Å and 0.016 Å for MACE-MP-0-small and MACE-MP-0-large, and 0.021 Å for M3GNet-MPF2021.2.8. The degradation relative to the inorganic-only MACE-MP-0 models is therefore small, especially for the large model (Shiota et al., 2024).

Bulk water provides an additional cross-domain test. In NVT molecular dynamics at 300 K, the O–O radial distribution function predicted by MACE-Osaka24-large with D3(BJ) is reported to be almost identical to that of MACE-MP-0-D3(BJ), indicating that the inorganic liquid-water behavior is effectively preserved. MACE-Osaka24-small-D3(BJ), by contrast, lies between MACE-MP-0 and MACE-OFF23 behavior, which suggests that finite model capacity interacts with dataset breadth in a visible way for condensed-phase dynamics (Shiota et al., 2024).

5. Coverage, limitations, and operating regime

MACE-Osaka24 is intended to be reliable across inorganic solids, organic molecules, water clusters, tripeptides, and at least some liquid-phase environments. Its training data include both equilibrium and non-equilibrium configurations, and its reported generalization extends to torsional scans, transition states, and bulk water structure. This makes it unusual among atomistic models in that molecular chemistry and crystalline materials are not treated as separate regimes but as one shared modeling domain (Shiota et al., 2024).

The principal explicit limitation discussed in the paper is the 4.5 Å cutoff radius. This becomes problematic for BCC crystals with large lattice constants, such as K, Rb, and Cs, where the physically relevant neighbor shells extend beyond the cutoff. In those cases, the local environment can appear artificially undercoordinated, and the model is not recommended. The authors therefore exclude some such systems from specific benchmark summaries. This is a reminder that “universal” in the MACE-Osaka24 context means chemically broad, not cutoff-free or fully long-range (Shiota et al., 2024).

A second limitation lies in TEA itself. The AEC step uses a single global scalar $N$7 for all systems and geometries. This is deliberately simple and robust, but it also means the alignment is globally linear rather than chemically adaptive. The paper explicitly notes that more flexible, possibly element-dependent or nonlinear alignment strategies are plausible future directions. This suggests that MACE-Osaka24 should be understood as a first-generation multi-domain foundation model enabled by TEA, rather than the endpoint of heterogeneous-data integration (Shiota et al., 2024).

6. Significance within the MACE ecosystem

MACE-Osaka24 matters for two distinct reasons. First, it demonstrates that one MACE network can cover both the MACE-MP-0 inorganic regime and the MACE-OFF23 organic regime with only modest compromise and, in some reaction benchmarks, with outright improvement. Second, it provides a practical recipe for democratizing universal MLIP construction: rather than requiring all contributors to recompute data at one reference level, TEA permits heterogeneous community datasets to be aligned and merged at low cost (Shiota et al., 2024).

In architectural terms, this development is a natural continuation of the broader MACE program. Earlier evaluations had already shown that MACE can achieve strong performance across amorphous carbon, universal materials modeling, small-molecule organic chemistry, large molecules, and liquid water, while retaining data efficiency and MD stability (Kovacs et al., 2023). MACE-Osaka24 extends that trajectory by addressing not only model architecture but also the data-integration problem that had previously separated molecular and crystalline foundation models into distinct silos.

A plausible implication is that future chemistry foundation models will depend at least as much on dataset harmonization as on neural-network design. In that sense, MACE-Osaka24 is both a model and a data-engineering statement: the universal potential is enabled not only by MACE’s equivariant message passing, but by the assertion that heterogeneous electronic-structure datasets can be placed onto a common energy scale with a controlled linear alignment. The public release of model weights, code, and the TEA-MPtrj/OFF23 dataset makes that claim operational rather than merely conceptual (Shiota et al., 2024).