Interatomic Potentials: A Framework for Generating Quantum-Accurate Material Models

Amit Samanta | 21-ERD-005

Project Overview

Starting with the pioneering work of Bernie Alder in 1950s, Lawrence Livermore National Laboratory has a long history of using atomic-scale molecular dynamics (MD) simulations both to study forefront scientific questions and to deliver critical material models to meet programmatic needs. The relentless growth in high-performance computing has greatly increased the range of important applications. However, a fundamental technical gap, i.e., the ability to quickly generate interatomic potentials for MD simulations, stands in the way of realizing the enormous power of these simulations. To fill this gap, we have developed a robust capability for high throughput generation of reliable and accurate interatomic potentials for metals and alloys at ambient and high-pressure conditions. The capability incorporates benefits of traditional physics-informed phenomenological models, the flexibility of modern machine learning models and emerging ideas from data science. The primary advantages of the interatomic potential development capability are: (a) the use of Gaussian basis functions as opposed to specific functional forms or orthogonal basis functions offers ample flexibility to accommodate diverse atomic environments and simplifies the optimization of free parameters; (b) the adoption of an iterative linear regression framework to minimize the number of free parameters allows for training potentials with considerably less training set data; (c) the nonlinear terms incorporate additional higher-order, many-body interactions that help in providing thermal stability while maintaining a reasonable number of coefficients. Our novel framework will enable challenging applications that are currently intractable due to the lack of accurate force laws, ranging from materials under high energy density (HED) regimes to the properties of complex additively manufactured materials.

Mission Impact 

The developed capability significantly reduces the time needed to generate reliable and accurate interatomic potentials. For example, in case of high-pressure iron (and separately for iron-nickel alloy), we were able to generate a training set in two days and train an ensemble of ten potentials in three days with very good agreement with quantum simulation results based on density functional theory (DFT) calculations for cold-curves, phonon dispersion curves, and enthalpy difference between solid and liquid at melt line. Since potentials are needed to develop material models, such as equation of state (EOS), phase diagrams, phase transition kinetics and strength models which are required for different programmatic efforts at Livermore, the developed capability is expected to play a critical role in developing better material models. Analysis of data from new HED experimental platforms is critical to achieving programmatic deliverables for Nuclear Weapons Stockpile Stewardship and the developed capability will allow for better understanding of data via atomic level simulations. In addition, it will aid in understanding unique properties of advanced multi-component alloys/materials (relevant to weapons physics) such as those prepared using additive manufacturing. This work is being leveraged in a new SI focused on assessing uncertainty estimation of fitted interatomic potentials, where generalized embedded atom method (GEAM) or GEAM's performance is being compared against other models and approaches, from an uncertainty quantification perspective.

Publications, Presentations, and Patents

Awards: LDRD team members have received SPOT awards from the Physics division for significant contributions to project deliverables (Haoyuan Shi, Jun Yang, Bajrang Sharma), mentoring high school students (Shashikant Kumar, Hong Sun, YingShi Teh, Amit Samanta) and for organizing the condensed matter seminar series (Amit Samanta).

Sun, Hong, and Amit Samanta. 2023. "Exploring structural transitions in grain boundaries of Nb using a generalized embedded atom interatomic potential." Computational Material Science 230:112497. LLNL-JRNL-849170.

Sharma, Bajrang, YingShi Teh, Babak Sadigh, Sebastien Hamel, Vasily Bulatov, Amit Samanta. 2023. "Development of an interatomic potential for the W-Ta system." Computational Material Science 230:112486. LLNL-JRNL-843995

Yang, Jun, Zhitao Chen, Hong Sun, Amit Samanta. 2023. "Graph-EAM: An interpretable and efficient graph neural network potential framework." Journal of Chemical Theory and Computation 19 2023.  LLNL-JRNL-846156

Zhu, Siya, Jibril Shittu, Aurelien Perron, Chirag Nataraj, Joel Berry, Joe McKeown, Axel van de Walle, and Amit Samanta, 2023. "Probing phase stability in CrMoNbV using cluster expansion method, calphad calculations and experiments." Acta Materialia 255:119062. LLNL-JRNL-845714

Kumar, Shashikant, Babak Sadigh, Vasily Bulatov, Sebastien Hamel, Brian Gallagher, John Klepeis, and Amit Samanta, 2022. "Accurate parametrization of the kinetic energy functional for calculations using exact exchange." Journal of Chemical Physics 156:024107. LLNL-JRNL-827314

Kumar, Shashikant, Edgar Landinez, Babak Sadigh, Vasily Bulatov, Sebastien Hamel, Brian Gallagher, John Klepeis, Amit Samanta. 2022. "Accurate parametrization of the kinetic energy functional." Journal of Chemical Physics 156:024110. DOI: 10.1063/5.0063629. LLNL-JRNL-828938

Hong Sun, Amit Samanta, Vincenzo Lordi, and Yayoi Takamura, "Exploring interface structure between perovskite oxides using evolutionary structure search and automated design of deep learning via neural architecture search" (Presentation, MRS Fall Meeting, Nov-Dec Boston, MA, 2022). LLNL-PRES-842752

Hong  Sun, Zhitao Chen, and Amit Samanta,"Interpretable physics-inspired Graph neural network force fields for atomistic modeling" (Presentation, MRS Spring Meeting, San Francisco, April 2023). LLNL-PRES-847125

Amit Samanta, "Novel methods to probe structure-property relationships at atomic scale" (Presentation, Workshop on Electronic and Photonic Materials, University of California Davis, September 2023). LLNL-PRES-854756