Session S50: Quantum-Accurate Atomistic Simulations at Extreme Scales: Recent Advances and New Challenges

A Universal Interatomic Potential for the Periodic Table

Mathematical graphs are a natural, universal representation for materials. In this talk, I will discuss the development of graph neural network (GNN) models as surrogate models for property predictions and interatomic potentials. Utilizing large federated databases such as the Materials Project, MatErial Graph Network (MEGNet) models can be trained to predict key properties such as the formation energy and band gap to sufficient accuracy for materials design and discovery. I will also discuss how some of the key limitations of GNN models, e.g., data hunger, can be addressed by incorporating global state variables. Finally, I will highlight our recent work on integrating traditional many-body formalisms into MEGNet models. The resulting M3GNet architecture can be used to train universal interatomic potentials that can work reliably across the entire periodic table. These advances enable ML-accelerated materials design across massive, diverse chemical spaces.   

Atomistic simulation of solid-solid phase transition from machine learning force fields

We present an efficient framework that combines machine learning potential (MLP) and advanced sampling techniques to investigate solid-solid phase transition. To achive this goal, we have developed a scalable MLP model to warrant an accurate interpolation of the energy surface where more than two solid phases coexist. In this presetation, two application examples will be discussed. We first combine the MLP with metadynamics simulation to investigate the phase transition of GaN under high pressure with different model sizes, in which we observe the sequential change of transition mechanism from collective modes to nucleation and growths. Such a mechanism change highlights the importance of statistical sampling with large system size. Secondly, we combine MLP with Monte Carlo simulation to elucidate the short range ordering on the NbMoTaW multi-principal element alloy (MPEA). The results show the strong attraction among Mo-Ta pairs forming the local ordered B2 structures. In addition, the property simulation results suggest that SRO increases the elastic constants and high-frequency phonon modes as well as introduces extra lattice friction of dislocation motion. This approach enables a rapid compositional screening and paves the way for computation-guided materials design of new MPEAs with better performance.

    

Modular and Scalable Solutions for Machine Learned Models for Material Science and Beyond

Model development that utilizes machine learning must define feature sets, model forms, and ultimately if said models are efficient to use for the desired accuracy. 

Specialized software to develop machine learned interatomic potentials utilized in MD is highlighted herein. The FitSNAP code has evolved quickly in the last few years to accept numerous descriptor sets, model forms (and associated regression techniques), all while ensuring portability into LAMMPS for efficient use. This talk will overview the user friendly FitSNAP code and its integration into the Exascale Computing Project EXAALT software stack with a focus on the challenges and advances made to tackle exascale sized training sets needed to construct robust and truly transferable interatomic potentials. Additionally, a span of applications will be highlighted that motivate a paradigm shift in how training data is curated for machine learned models. A pair of new methods for training set generation that is applicable beyond interatomic potentials will be demonstrated to this end.    

Training Machine learned Interatomic Potentials to EXAFS Data for Simulations Under Extreme Conditions

Machine Learned (ML) interatomic potentials are an extremely popular tool for running nearly quantum accurate molecular dynamics (MD) simulations at large scales. Traditionally, ML potentials are trained to energies and forces derived from quantum mechanical simulations. In this paradigm, experimental measurements are relegated to the role of testing data. While quantum mechanical data has proven successful for training ML potentials, there are cases where it is prohibitively expensive to obtain or where quantum models for a system are inaccurate and produce incorrect predictions. With the increasing prevalence of high-resolution, high repetition rate experiments occurring at modern laser and synchrotron facilities, the ability to train ML interatomic potentials directly to atomistic experimental probes is becoming increasingly important. Here we will describe a general-purpose training procedure for refining ML interatomic potentials on experimental data, only requiring a forward model for comparison to experiment. This procedure will first be demonstrated with processed radial distribution function data, as extracted from x-ray diffraction, to refine an aluminum ML interatomic potential. As a second example of this procedure, an ML potential for zinc, originally trained to DFT data, is refined by direct comparison to EXAFS (Extended X-ray Absorption Fine Structure) spectra taken at elevated temperatures and pressures. Both models are then tested against out of sample experimental data such as diffusion constants, elastic properties, and others. Finally, the potentials are used run large scale MD simulations to interpolate the experimental measurements for new pressure and temperature regimes.   

Towards Automated and Robust Atomic Cluster Expansion Models

The Atomic Cluster Expansion (ACE) is a general and systematic linear model for representing interatomic potentials and force fields. In this presentation, I will present our latest efforts to automate the selection of training sets, feature selection, and parameter estimation via a combination of uncertainty quantification, a variant on active-learning, robust regression techniques and injecting modelling insight (geometric priors) into the model formulation. Show-case applications include fully automated generation of phase diagrams for multi-component alloys. Time permitting I will discuss early attempts to explain the impressive generalisation capabilities of ACE (and MLIPs in general).