Machine Learning Force Fields

The development of machine learning models has led to an abundance of datasets containing quantum mechanical (QM) calculations for molecular and material systems. However, traditional training methods for machine learning models are unable to leverage the plethora of data available as they require that each dataset be generated using the same QM method. Taking machine learning interatomic potentials (MLIPs) as an example, we show that meta-learning techniques, a recent advancement from the machine learning community, can be used to fit multiple levels of QM theory in the same training process.

Graph Neural Networks (GNNs), especially equivariant message-passing neural networks (MPNNs), have emerged as powerful architectures for machine learning force fields. However, MPNNs face challenges when modeling non-local interactions in systems such as long conjugated molecules, due to oversmoothing and oversquashing. To address these concerns, we introduce Matrix Function Neural Networks (MFNs), a novel architecture that parameterizes non-local interactions through learnable equivariant matrix functions. Employing resolvent expansions offers a straightforward implementation and the potential for linear scaling with system size. The MFN architecture achieves state of-the-art performance in standard graph benchmarks, and is able to capture intricate non-local interactions in quantum systems such as long conjugated molecules, paving the way to new state-of-the-art machine learning force fields.

Machine-learning interatomic potentials (MLIAPs) make it feasible to aim for both accuracy and transferability, something that the earlier generations of potentials struggled to achieve. However, given their flexibility, these potentials often fail at extrapolating to properties beyond the training data, which makes the quality of the training set one of the determining factors in their performance. Training datasets typically consist of DFT energies and forces of relatively small systems, traditionally selected manually or randomly from subsets of the configuration space of interest. The need for human intervention in the curation of training sets makes their generation labor-intensive and time-consuming. Here, we present a generalization of a previously developed method based on the automated maximization of the information entropy of the descriptor distribution. The diversity of the entropy-optimized training dataset is compared to several other datasets from the literature. This approach is demonstrated for several elements, which produced potentials that are very robust in a broad range of conditions, highlighting the desirable characteristics of an optimal training data, irrespective of the material chemistry.

Data-driven interatomic potentials, like atomic cluster expansion (ACE), are revolutionizing material simulations. These models achieve quantum-mechanical accuracy in approximating the potential energy surface, enabling researchers to explore previously inaccessible time and length scales. However, creating a comprehensive training database specific to the material remains a challenge. In this talk, we propose an information-theoretic approach to streamline the design of force fields for material-specific applications using linear ACE potentials. This method will be shown to create accurate potentials for complex alloys like AlSi10Mg, Ti6Al4V and MoNbTaW, allowing us to investigate phase transitions (solid-solid, solid-liquid) and predict various metallurgical phenomena, including precipitation hardening and the discovery of dual phases.

Physics-based, atom-centered machine learning (ML) representations have been instrumental to the effective integration of ML within the atomistic simulation community. Many of these representations build off the idea of atoms as having spherical, or isotropic, interactions. In many communities, there is often a need to represent groups of atoms, either to increase the computational efficiency of simulation via coarse-graining or to understand molecular influences on system behavior. In such cases, atom-centered representations will have limited utility, as groups of atoms may not be well-approximated as spheres. In this work, we extend the popular Smooth Overlap of Atomic Positions (SOAP) ML representation for systems consisting of non-spherical anisotropic particles or clusters of atoms. We show the power of this anisotropic extension of SOAP, which we deem AniSOAP, in accurately characterizing liquid crystal systems and predicting the energetics of Gay-Berne ellipsoids and coarse-grained benzene crystals. With our study of these prototypical anisotropic systems, we derive fundamental insights into how molecular shape influences mesoscale behavior and explain how to reincorporate important atom-atom interactions typically not captured by coarse-grained models. Moving forward, we propose AniSOAP as a flexible, unified framework for coarse-graining in complex, multiscale simulation

Recently, Machine Learning Interatomic Potentials have emerged as versatile surrogate models capable of accurately reproducing ab initio potential energy surfaces. However, most of their applications have been targeted at near-equilibrium closed-shell structures. In this talk I  introduce a transferable MACE interatomic potential that is applicable to open- and closed-shell drug-like molecules containing hydrogen, carbon, and oxygen atoms. Including an accurate description of radical species extends the scope of possible applications to bond dissociation energy (BDE) prediction, for example, in the context of cytochrome P450  metabolism. I will discuss the model’s performance and compare it to that of alternative methods for BDE prediction. Additionally, I will provide an outlook for extending the model’s scope to modelling the full C-H abstraction pathways.

While first principles molecular simulations have been pivotal for material science in general, numerous phenomena including nanoparticle formation require a drastic reduction of the computational costs. With this objective in mind, machine-learning interaction potentials have been recently developed for lots of different applications and numerous high-impact studies have already demonstrated the great efficiency of these techniques. After this first breakthrough, the current effort is on investigating the limitations associated with those techniques and on developing ways for obtaining similar accuracy with simpler models. In this contribution, we will present Physical LassoLars Interaction Potential (PLIP) which is a method combining a linear formulation of the potential with a constrained regression scheme. We will then demonstrate the ability of PLIP to generate an accurate model for zinc oxide interactions and provide a thorough comparison with neural network potentials in terms of transferability. Next, we will describe how PLIP can be extended to account for long-range interactions. The final part of this contribution will be dedicated to a study of crystal nucleation in nanoparticles of zinc oxide where we will show that different nucleation pathways are competing depending on the studied degree of supercooling.