Assessing zero-shot generalisation behaviour in graph-neural-network interatomic potentials

TL;DR

This work studies GO-MACE-23, a model designed for the extended covalent network of graphene oxide, and quantifies its zero-shot performance for small, isolated molecules outside its direct scope, as well as for examples of chemical reactions.

Abstract

With the rapidly growing availability of machine-learned interatomic potential (MLIP) models for chemistry, much current research focuses on the development of generally applicable and ``foundational'' MLIPs. An important question in this context is whether, and how well, such models can transfer from one application domain to another. Here, we assess this transferability for an MLIP model at the interface of materials and molecular chemistry. Specifically, we study GO-MACE-23, a model designed for the extended covalent network of graphene oxide, and quantify its zero-shot performance for small, isolated molecules and chemical reactions outside its direct scope--in direct comparison with a state-of-the-art model which has been trained in-domain. Our work provides quantitative insight into the transfer and generalisation ability of graph-neural-network potentials and, more generally, makes a step towards the more widespread applicability of MLIPs in chemistry.

Figures (7)

• Figure 1: Visualising the structural and chemical space explored in the present study. We show a two-dimensional embedding of the MACE descriptor trained on the GO dataset elmachachi_accelerated_2024, using principal component analysis. The points of the map correspond to the training set of GO-MACE-23 (blue), molecules containing C, H, and O atoms, representing $\approx 5$% of the SPICE (version 1) dataset eastman_spice_2023 (red), configurations from rMD17 trajectories christensen_role_2020 (purple), a series of fullerenes with sizes ranging between 20 and 100 (magenta), five molecules encapsulated in C60 fullerene cages (yellow), and the same molecules in vacuum (black crosses).
• Figure 2: Energy and force errors on six trajectories from the revised MD17 dataset using GO-MACE-23. The bars represent the RMSE of quantities between GO-MACE-23 predictions and rMD17 labels. The dashed area represents the errors between the DFT levels of theory used to label the GO dataset and the rMD17 dataset. The dashed line is the internal validation error of GO-MACE-23.
• Figure 3: (a) Visualisation of a toluene molecule obtained using OVITO stukowski_visualization_2010. Red- and blue-coloured atoms are carbon atoms part of the aromatic ring and the attached methyl group, respectively. (b) Force components parity plot of the DFT-computed and GO-MACE-23-predicted forces for the carbon atoms labelled red and blue in panel (a). (c) Force parity plot of the sum of forces of the red- and blue- labelled carbon atoms.
• Figure 4: Molecular vibrational spectra computed with MLIPs ( solid lines) and DFT ("QM", dashed lines) for GO-MACE-23-relaxed naphthalene, toluene, and malonaldehyde molecules. The upper row characterises the out-of-domain performance of GO-MACE-23 ( red). The lower row shows the performance of SOTA MLIPs for molecules, viz. MACE-OFFkovacs_mace-off23_2023 ( dark and light blue). Note that the DFT data have been computed at the level corresponding to the training data of the respective MLIP model; the DFT data in the upper and lower rows are therefore slightly different.
• Figure 5: Evolution of the per-atom energy of fullerenes, obtained from Ref. barnard_fullerene_2023, of sizes between 20 and 100 atoms computed with GO-MACE-23 and its corresponding DFT level of theory ( red), and MACE-OFF and their corresponding DFT level of theory ( dark and light blue). Similar to Fig. \ref{['fig:phonons']}, lines represent the ML predictions, and the dashed lines represent the QM reference calculations. All energies are referenced to C60. The lower panel describes the difference between energies computed with ML and QM, and expressed per atom. The rendered images show three fullerenes: C20, C60, and C100.