Hello, A yearly allocation for the LCRC cluster has been requested with the following updated information: Submitter/PI: John J. Low Project Name: perfect Division: EGS, CLS, LCF, APS, MSD Project title: The Perfect Thermodynamics of Imperfect Materials Associated funding: LDRD Other Systems: ALCF/INCITE Science: Abstract: The goal of the project is to significantly improve the understanding and prediction of thermodynamic stability/metastability of “imperfect” (e.g., highly defective, non-stoichiometric, or doped) oxide material phases, via innovative theory (i.e., embedded uncertainty), advanced experiments at APS, and “intelligent” software (i.e., able to learn and quickly solve new problems). We envision building the knowledge and capabilities that will allow, over the next decade, the prediction of thermodynamic properties of imperfect materials, with impact on materials design, synthesis, and smart manufacturing. Furthermore, we expect this methodology to accelerate the development of the materials genome and the next generation computers. Science: We focus on high-k dielectric materials for complementary metal-oxide-semiconductor (CMOS), which are of particular importance for creating Dynamic Memory Allocation (DRAM) devices. Many CMOS properties strongly depend on material defects such as vacancies, interstitials, defect clusters that occur during synthesis, and thermal treatment. The inclusion of other chemical elements (e.g., dopants) in CMOS can significantly change physical properties such as thermal conductivity, electrical conductivity, and magnetism. Far from representing a drawback, the understanding imperfections, including defects and dopant, leads to prediction and control of properties and the design of post-CMOS materials for the next generation of electronic devices. There is, however, a challenge: the thermodynamic stability of CMOS phases – a crucial element for any application – is hard to evaluate, especially at high temperatures, close to the phase transformations or the melting point. In this project, we develop the theoretical, experimental, and computational knowledge and tools for predicting the thermodynamic and kinetic stability/metastability of defected CMOS phases at temperatures ranging from 0K to melting. As a proof of principle, we will study the high k dielectric in complementary metal–oxide–semiconductor (CMOS), using as a prototype hafnia, HfO2. Hafnia has unusual electronic and magnetic properties that make it attractive to both science and technology. One of the most critical applications is based on hafnia’s high-k dielectric properties, which are of particular importance for creating Dynamic Memory Allocation (DRAM) devices. In 2007, Intel decided to replace silicon oxide with hafnia as a gate insulator for field-effect transistors. Current and future technology (sub 14nm) are using “hafnia based” materials (including additional elements). Many CMOS properties strongly depend on materials defects such as vacancies, interstitials, defect clusters (Figure 1), and how the material was synthesized and pretreated. We will extend this studies to defects in the tetragonal, cubic and molten HfO2. Throughout the project, we will compare the computational results to available experimental and theoretical data. We will focus on oxygen vacancies, hydrogen interstitials, and nitrogen dopants, which are common defects in HfO2 high k-dielectrics. Additional work will model the crystallization of amorphous hafnia. Project description: In a preliminary study, we have calculated enthalpies of several Hf-O phases as a function of temperature using ab initio molecular dynamics (AIMD). The calculations involved a few dozen trajectories spanning a temperature range from 0K to a 4000K (above the melting temperature of hafnia) for each model of the perfect and defective phases. Each trajectory involves 10 picoseconds of dynamics (10,000 time steps) or 14 hours on 16 Broadwell nodes (8064 core-hours). Calculating two dozen trajectories requires 193,536 core-hours. We have completed trajectories for perfect and defective monoclinic, tetragonal and cubic polymorphs during bebop’s early user period. We still need to complete the calculations for hydrogen interstitials, nitrogen interstitials, and the molten HfO2 phase. The results of this work will be used to develop a force-field Hafnia using Machine Learning by Y. Ghadar and A. Vazquez Mayagoitia. We estimate that the force-field develop ment work will require 2,000,000 core-hours to complete. Hafnia in semiconductors is a thin (10-100 nm) amorphous layer between layers of Si. The synthesis process determines the structure of this nanolayer. Typically, atomic layer deposition or chemical vapor deposition is used to fabricate these monolayers. We will follow the procedure proposed by Shirazi and Elliot [J. Comp. Chem. 35(2014)244] and use Kinetic Monte Carlo (KMC) to predict the structure of the layer of amorphous HfO2 grown on Si during ALD or CVD. The predicted reaction rates, needed as input in KMC, will be derived from density functional calculations. To solve the KMC equations, we will use Eon [Henkelman and Jonsson, JCP 115(2001)9 657] and SPPARKS [Plimpton et al. JCP 128(2008)205101]. The KMC simulations will require 250,000 core-hours. A critical parameter is the crystallization temperature of the amorphous layer of hafnia. Crystallization is detrimental to a dielectric property because it can lead to current leakage. To avoid leakage, the processing temperature of the DRAM chip must be less than the crystallization temperature. Since the amorphous phase is metastable, the standard process of estimating free energies from energies and phonon bands to predict phase transitions is not adequate. We propose to use Forward Flux Sampling (FFS) [Allen et al. J. Phys.: Condens. Matter. 21(2009)463102] to model the crystallization process, predict rates, and predict the crystallization temperature. We will use reactive force-fields in LAMMPS to model this process and will develop LAMMPS scripts to conduct FFS through collaboration with Art Voter (LANL). We estimate that this study will require 100,000 core-hours. All data resulting from theory and calculations will be evaluated using Bayesian Analysis (BA) to determine the optimal models (i.e., mathematical expressions) that relate the free energy of the defected oxide phases to temperature and defect concentration (first and second year). To this end, we will use a Non-Naive Bayesian Network machine learning algorithm and software to process that will analyze the training and validation data sets in an unsupervised mode. The data generated by the DFT and KMC computations described above will enrich the diversity of the samples and reduce the final error in our estimations. The computationally intensive calculations will improve the evaluation of the stability and metastability of phases and will add uncertainty intervals, using a methodology developed by the PI and co-authors (M. Stan and B. Reardon., CALPHAD, 27 (2003) 319). We estimate that 100,000 core-hours will be required for these calculations. We will explore the scalability of alternative electronic structure programs as part of this project. Since a majority of the CPU time consumed in this project is by electronic structure programs, we evaluated several DFT codes such as CP2K and VASP to select the most efficient and scalable electronic structure program. The evaluation shows that VASP6 with multiple k points is the most accurate for smaller unit cells and exhibits competitive performance. OpenAtom is a new high-performance electronic structure program which claims good scaling up to ten thousand cores on BG/Q and blue-waters. This scaling would improve throughput by a factor of ten, which would be extremely useful in AIMD studies. We plan to use this OpenAtom to model molten and amorphous HfO2. This study will require 1,000,000 core hours. Industry partnership: Project URL: Current FY Hours Used: undetermined amount New FY Requested allocation: 3500000 Q1: 875000 Q2: 875000 Q3: 875000 Q4: 875000 Justification: LAMMPS is a high-performance code which has been optimized to run on Broadwell, NVidia GPUs, and Knights Landing processors. Molecular dynamics for thousands of atoms will be used to model crystallization which will scale well to thousands or cores. See http://lammps.sandia.gov/bench.html for more details. Version 6 of VASP is a hybrid MPI/open executable which can scale to a thousand cores on the Broadwell processors on bebop and Sandy Bridges processors on blues for unit cells containing ~100 atoms of HfO2. We observe good scaling for up to 8 KNL nodes with performance competitive with the Broadwell nodes. VASP has been our main engine for generating thermodynamic data from first principles. See https://cug.org/proceedings/cug2017_proceedings/includes/files/pap134s2-file... for details on the new version of VASP. OpenAtom is a high-performance electronic structure program which uses Charm++ and has an improved FFT algorithm. It claims good scaling up to ten thousand cores on BG/Q and Blue Waters. See http://charm.cs.illinois.edu/OpenAtom/research.shtml for details. Eon and SPPARKS are both designed to run parallel. Eon can run multiple LAMMPS jobs simultaneously while SPPARKS can run multiple KMC jobs and has demonstrated good parallel scaling to a thousand cores or more. See http://henkelmanlab.org/pubs/chill14_055002.pdf for details. Storage requirements: Thank You, The LCRC Accounts System