Collaborative Research: Reliable Materials Simulation based on the Knowledgebase of Interatomic Models (KIM)

NONTECHNICAL SUMMARY

This award supports OpenKIM, a cyberinfrastructure component of the research community that uses computer simulations of atoms based on Newton's Laws and models for the interaction between atoms, to attack problems in materials science, engineering, and physics, and to enable the discovery of new materials, design new devices, to advance the understanding of materials-related phenomena, and much more. Recent years have seen significant advancement in the areas of materials knowledge, discovery, and manufacturing methodologies. This includes, for example, the development of graphene (a single atomic layer of carbon atoms, which has exceptional mechanical, thermal, and electrical properties) and the related class of two-dimensional materials with unprecedented material properties now being extensively studied by scientists and engineers. Another example is the advent of three-dimensional printing techniques that allow engineers to design new materials from the ground up that can be tailor-made for their specific application. Computer simulation of materials at the atomic-scale is one of the key enabling technologies driving the current materials revolution. Although the most accurate atomic-scale simulations employ the equations of quantum mechanics, such computations take so long to complete, even on today's powerful computers, that practically they are limited to a few thousands of atoms. This is simply not enough for the study of materials properties, which requires the simulation of interactions between millions and even billions of atoms. Thus, materials researchers rely on faster more approximate equations, known as interatomic models (IMs), to describe atomic interactions. These models are fast, but typically they are only accurate for a restricted range of material properties. This limited range of applicability necessitates the creation of many IMs, even for a single material such as silicon. Organizing, sharing, and evaluating the range of applicability of these IMs has been a long-standing challenge for the materials research community. In most cases researchers have no way of knowing which IM is suitable for their particular application. Further, the proliferation of IMs, often designed to work only with specific simulation programs, makes it difficult to share and exchange IMs, and to reproduce other researchers' work, which is how science evolves and self corrects.

The Knowledgebase of Interatomic Models (KIM) is a project that is working to solve these challenges. To date, the KIM project has developed an online framework at https://openkim.org to address the issues of IM provenance, selection, and portability. IMs archived on this website are exhaustively tested and can be used in plug-and-play fashion in a variety of major simulation codes that conform to a standard developed as part of the KIM project. The development activity of the current project will extend the KIM framework by broadening the number and types of supported IMs, and will add new capabilities and educational resources that will make it easy for researchers to integrate the IMs and materials data available on openkim.org into their daily research workflow. Further, emerging techniques in information topology and machine learning will be applied to study and quantify the inherent uncertainty in predictions made by IMs, and to assist materials researchers to select the best IM for their application. Together the development, educational, and research activities of this project are expected to significantly increase the userbase and broader impact of the KIM project.

TECHNICAL SUMMARY

This award supports OpenKIM, a community Knowledgebase of Interatomic Models (KIM) for simulation. KIM is a project for normalizing the use of IMs in molecular simulations of materials. An IM, often referred to as a 'potential' or 'force field,' is an approximate method for computing the energy and its derivatives for an atomic configuration. This project addresses both traditional 'physics-based' IMs and the new class of 'data-driven' IMs introduced in recent years. In a sustained effort, the KIM project has developed a systematic framework to address the IM provenance, selection, and portability problems faced by materials researchers. Before KIM, these challenges were the cause of significant inefficiencies and inaccuracies in the research pipeline. Today, an IM available on openkim.org is subjected to a rigorous set of 'Verification Checks' that aim to ensure that its implementation conforms to a high software-engineering standard, and to an extensive set of 'Tests,' each of which computes a well-defined material property for assessing the IM's accuracy. A researcher can come to openkim.org and explore the predictions of KIM Models in comparison with experimental or quantum 'Reference Data' to select a suitable IM for their application. The current project is aimed at extending KIM to become an integral component of the workflow of researchers engaged in molecular simulation to make their work more efficient and their results more reliable and reproducible. To achieve this vision, the Principal Investigators (PIs) will pursue the following program of cyberinfrastructure R&D and basic research related to IM usage and science. The cyberinfrastructure R&D will include extensions to KIM standards to support additional common IM features (such as long-range fields) and added support for IMs having cutting-edge features that cannot yet be standardized. Further, KIM will be integrated into existing simulation tools so that researchers may query and retrieve data archived on openkim.org as part of their daily workflow. This approach reduces errors, ensures reproducibility, uses a standard tested method (embodied in a KIM Test) to obtain the desired property, and firmly integrates the KIM framework into the workflow of computational materials researchers. The basic research component of the project includes three research thrusts requiring advances to enhance the reliability of molecular simulations: (1) IM Uncertainty: The PIs will use ideas from information topology and differential geometry to automatically generate IM ensembles for obtaining estimates of the inherent uncertainty of the IM. (2) IM Transferability: The PIs plan to adapt a multi-task machine learning approach to predict an IM's accuracy for different applications. This will lead to a rigorous, objective criterion to assist researchers with IM selection. (3) IM Heuristics: By mining IM predictions and Reference Data archived on openkim.org, it is possible to identify correlations similar to empirical heuristics such as Vegard's rule and connections between microscopic properties and macroscopic features. Detection of such heuristics will provide insights into the limitations of IMs, help design optimal training sets, and lead to better understanding of the properties of IMs generally. In terms of broader impacts, the scope of the KIM project is unusually large - far beyond materials science - due to the prevalence of molecular simulations across the physical sciences from microbiology to geology. The project aims to maximize its impact by (1) expanding the KIM user base, (2) engaging the materials research community directly and through targeted research and educational efforts, and (3) developing new relationships and collaborations with other materials modeling cyberinfrastructures and organizations.

This award is jointly supported by the Division of Materials Research in the Directorate for Mathematical and Physical Sciences and the Civil, Mechanical and Manufacturing Innovation Division in the Engineering Directorate.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.