Choosing a foundation potential

Foundation potentials (FPs), i.e., universal machine learning interatomic potentials with periodic-table-wide coverage, are now proliferating. New pre-trained FPs seem to appear almost daily, making hyped-up claims about performance. For those not deep in the field, it’s becoming increasingly difficult to separate signal from noise.

Here’s my expert, very opinionated take on how to choose an FP that actually fits your needs.

1️⃣ Dataset quality and size drive everything. Architecture is a secondary consideration.

If you’re working with inorganic crystals or condensed matter, there are really only two datasets that matter today: OMat24 and MatPES (yes, I’m involved with the latter, but I’m being objective here).

OMat24 wins on sheer size and broad coverage near equilibrium, but its convergence criteria based on MPRelax parameters are a bit loose for true PES quality, i.e., the kind that people use to train plain MLIPs. But at least OMat24 is constructed from static calculations, not merely sampled from intermediate steps in a relaxation trajectory or MD, such as MPTrj.

MatPES is smaller, but its parameters are modernized for PES accuracy - the latest pseudopotentials, strict energy and force convergence, etc. Our experience is that the predictions are much more reliable, and it is much easier to work with. It is the only option if you want r2SCAN PES information, which for many systems, is far superior to PBE.

Pick either one. The community should retire MPTrj — it was never meant for MLIP training and was just the best available dataset in the early days of FPs.

2️⃣ Don’t mix datasets unless you really know what you’re doing.

I’m still amazed how often people blend OMat24, Alexandria, and MPTrj thinking it improves generality.

It doesn’t.

Mixing a high-quality dataset with a low-quality one yields a low-quality model.

3️⃣ Choose architecture efficiency over novelty.

Unless you’re after new physics beyond PES quantities (like charges or magnetism), focus on architectures that give the best accuracy with the fewest parameters and lowest inference cost. At this point, the major FP architectures perform similarly in terms of accuracy. The real differentiator is computational efficiency. Some of the more models are more than 1000× slower at inference and can barely handle systems beyond 100 atoms. This defeats the entire purpose of a FP: being a practical surrogate for DFT.

If you just want to play around with different FPs to check their performance on a property that matters to you, MatCalc is a good choice to get this done with a minimal amount of coding.