MACE-MP-0 - A Foundation Model for Materials Chemistry

MACE-MP-0 is a neural network potential (NNP) trained to reproduce the results of DFT calculations run using the PBE functional, from Gábor Czányi's group at Cambridge. Since this was released late in 2023, the paper has been incredibly influential in the field: it already has 85 citations, and virtually everyone I know in atomistic simulation has read and discussed the results. Here's what's so cool about MACE-MP-0, and why we're very excited to bring this model to Rowan:

There Aren't Forcefields For Materials Science

In contrast to NNPs focused on organic chemistry or biomolecular chemistry, like AIMNet2, MACE-OFF23, and plenty of others, MACE-MP-0 is focused on reproducing periodic DFT calculations for bulk materials of all kinds. Effective low-cost methods in this area are rare: the forcefields commonly used to model bioorganic systems don't extend to these systems, and forcefields that cover the entire periodic table (like UFF). Low-cost semiempirical quantum methods like xTB have become useful, but support for periodic systems remains limited and the speed of these methods is often insufficient.

For accurate simulations of materials, periodic DFT thus is conventionally the only option. Unofrtunately, periodic DFT is infamously compute-intensive. This is where MACE-MP-0 comes in—it's a reasonably accurate model, it works on lots of systems, and computations with MACE-MP-0 are many times faster than DFT calculations.

MACE-MP-0: A Foundation Model?

The authors demonstrate that MACE-MP-0 shows tremendous success at modeling a vast variety of systems and processes: CO₂ binding in MOFs, NaCl dissolution in water, melt-quenched amorphous carbon structures, cerium oxide nanoparticules, sulfur polymerization, combustion of mixtures of hydrogen and oxygen, Jahn-Teller distortions in LiNiO₂, etc. The model is trained over virtually the entire periodic table, making it difficult to find systems that are out-of-distribution. For some example heterogenous catalysis systems (CO oxidation on Cu and CO₂ to methanol conversion on In2O3), MACE-MP-0 even produces transition states in roughly the correct places, although the overall energies are quite wrong.

There are some necessary caveats. MACE-MP-0 is trained against PBE data, so there are cases in which this causes it to give incorrect descriptions: for instance, MACE-MP-0 closely matches the PBE radial distribution function for liquid water, but PBE itself doesn't do a particularly good job modeling the structure of water.

Similarly, MACE-MP-0 seems to struggle when evaluating intermolecular interactions in organic systems:

• Densities of ethanol–water mixtures are qualitatively incorrect
• Volume scans of ionic liquids give poor results
• The X23 lattice energy results are okay, but not great
• And the free-energy surface for the alanine tripeptide is atrocious

But these are tasks which organic chemistry-specific models like AIMNet2 are quite good at! MACE-MP-0 seems to excel in systems that aren't able to be described by existing organic chemistry models, which makes it incredibly useful even if it's not able to solve all problems in all areas of chemistry.

When and How To Use MACE-MP-0

MACE-MP-0 won't replace DFT. The scope of systems that MACE-MP-0 aspires model is massive, and significant advances in dataset quality and network architecture will be needed to fulfill these aspirations. The PBE density functional used for training data is quite inaccurate for barrier heights and many intermolecular interactions, so even a model that perfectly learns the training data will still suffer from limited accuracy in these areas.

Even with these caveats, MACE-MP-0 is an incredible advance for computational materials science—it's many times faster than conventional methods and gives useful accuracy on a wide variety of architectures. If you want to try running a calculation with MACE-MP-0 right now, create an account on Rowan!