Algorithm Development in Materials Science and Engineering: Electronic Scale Calculations and Machine Learning
Sponsored by: TMS Materials Processing and Manufacturing Division, TMS: Integrated Computational Materials Engineering Committee, TMS: Phase Transformations Committee, TMS: Computational Materials Science and Engineering Committee
Program Organizers: Mohsen Asle Zaeem, Colorado School of Mines; Garritt Tucker, Colorado School of Mines; Charudatta Phatak, Argonne National Laboratory; Bryan Wong, University of California, Riverside; Mikhail Mendelev, NASA ARC; Bryce Meredig, Travertine Labs LLC; Ebrahim Asadi, University of Memphis; Francesca Tavazza, National Institute of Standards and Technology
Monday 8:00 AM
February 24, 2020
Room: 31C
Location: San Diego Convention Ctr
Session Chair: Mohsen Asle Zaeem, Colorado School of Mines; Bryan Wong, University of California, Riverside
8:00 AM Introductory Comments Mohsen Asle Zaeem
8:05 AM Invited
High-throughput Computational Design of Organic-inorganic Hybrid Halide Semiconductors Beyond Perovskites: Kesong Yang1; Yuheng Li1; 1University of California, San Diego
Organic–inorganic lead halide perovskites show great promise in optoelectronic applications such as light-emitting diodes and solar energy conversion. However, the poor stability and toxicity of lead halide perovskites severely limit their large-scale applications. Here we show a high-throughput design of lead-free hybrid halide semiconductors with robust materials stability and desired material properties beyond perovskites. On the basis of 24 prototype structures that include perovskite and non-perovskite structures and several typical organic cations, a comprehensive quantum materials repository that contains 4507 hypothetical hybrid compounds was built using large-scale first-principles calculations. After a high-throughput screening of this repository, we have rapidly identified 23 candidates for light-emitting diodes and 13 candidates for solar cell. Our work demonstrates a new avenue to design novel organic–inorganic functional materials by exploring a great variety of prototype structures. It is important to highlight that this approach is transformative to the discovery of other types of functional materials.
8:35 AM Invited
Stochastic Exchange for Efficient Long-range-hybrid-DFT for Thousands of Electrons and More: Daniel Neuhauser1; 1UCLA
Stochastic Quantum Chemistry (SQC) is a new paradigm we developed for electronic structure and dynamics, that rewrites traditional quantum chemistry as stochastic averages, avoiding the steep power law scaling of traditional methods. As an example I will discuss Stochastic Exchange (sX) for long-range hybrid DFT. Unlike local functionals, long-range hybrids are excellent for charged and polarized systems. The computational bottleneck of the long-range exchange is avoided in sX by a stochastic resolution of the identity, where both the Coulomb potential and the density matrix are each replaced by a direct-product of stochastic vectors. The method is very efficient and requires proportionally less effort for larger systems, so for systems with thousands of electrons it only doubles the DFT computation time compared with local methods. No density-matrix locality is assumed.
9:05 AM
Machine Learning Approaches for Improving Density Functional Tight Binding Models of Reactive Materials: Application to Astrobiolgical Materials and Surface Chemistry: Nir Goldman1; 1Lawrence Livermore National Laboratory
Density Functional Tight Binding (DFTB) methods can accurately probe chemical reactivity at nanosecond timescales but can be challenging to parameterize for each system of interest due to the different bonding types that can occur. Here, we have created a machine learning-based approach for determining the DFTB repulsive energy which is both rapidly optimized, systematically improvable, and highly transferable. Our method leverages the Chebyshev Interaction Model for Efficient Simulation (ChIMES), a reactive many-body molecular dynamics force field where interactions are represented by linear combinations of Chebyshev polynomials. We have created ChIMES/DFTB models for a wide variety of systems, including potential hydrogen storage materials and impacting astrophysical ices and the synthesis of life-building compounds. Our approach is easy to implement and can yield accurate DFTB models for a number of challenging materials and conditions where chemical properties can be difficult to model with standard quantum approaches alone.
9:25 AM
Uncertainty Quantification for Machine Learning Methods Applied to Material Properties: Kamal Choudhary1; Francesca Tavazza2; 1University of Maryland(National Institute of Standards and Technology; 2Umcp/Nist
Next generation material discovery and characterization is heavily dependent on the use of machine learning (ML) approaches. The key ingredients in most ML models are well curated data, descriptors and appropriate algorithms. As ML in materials become more popular, it is essential to quantify uncertainty in terms of descriptors and ML algorithms. In this work, we compare several descriptors and ML algorithms for DFT generated data for molecules, 3D and 2D solid materials. Some of the investigated material properties are formation energy, bandgap, refractive index, elastic constants, topological spillage, and exfoliation energy.
9:45 AM Break
10:00 AM Invited
Unraveling Exciton Dynamics in 2D Van der Waals Heterostructures: Junyi Liu1; Xu Zhang1; Gang Lu1; 1California State University Northridge
2D vdW heterostructures offer a fascinating platform to pursue fundamental science and device applications. Owing to reduced dielectric screening, excitonic effect is prominent in 2D materials. Recent experiments have revealed ultrafast charge transfer dynamics in transition metal dichalcogenides heterostructures, which is surprising giving the strong electron-hole binding of the excitons. Most theoretical calculations were performed based on local and semi-local exchange-correlation functionals, thus cannot capture the excitonic effect accurately. I will introduce a first-principles method that combines non-adiabatic molecular dynamics with linear-response time-dependent density functional theory. The method is formulated in planewave bases and pseudopotentials with range-separated hybrid functionals. As a result, the method can capture the excitonic effect accurately in extended 2D heterostructures. Using this method, we can shed light on the ultrafast charge transfer dynamics in vdW heterostructures and elucidate the role of “hot” excitons in promoting ultrafast charge transfer despite the momentum mismatch in twisted heterostructures.
10:30 AM Invited
Isolated Dislocation Core Energy from First Principles Energy Density Method: Yang Dan1; Dallas Trinkle1; 1University of Illinois at Urbana-Champaign
The energy per length of a dislocation controls a variety of phenomena in plastic deformation: the bowing of dislocations, the formation of kinks and jogs, and even provides a driving force for dislocation recovery. The line energy is the sum of two contributions: a long-range contribution that can be computed using linear elasticity theory and a "core" energy that can only be treated accurately with atomistic scale modeling. First principles calculations have been used to study dislocation geometry for the past two decades, but boundary conditions--either periodic, fixed, or flexible--have made the computation of the core energy of an isolated dislocation difficult. Recent advances in the first principles energy density method allow the spatial partitioning of energy, which in turn, permits new analysis of the dislocation core and definition of the core energy. We will showcase this approach with basal and prismatic a-type screw dislocation cores in magnesium.
11:00 AM
A First Principles Multi-cell Monte Carlo Method for Phase Prediction: You Rao1; Changning Niu2; Wolfgang Windl1; Maryam Ghazisaeidi1; 1The Ohio State University; 2QuesTek Innovations LLC
The importance of phase prediction lies not only in the understanding of basic thermodynamics, but also in the development of new materials. This has been a challenging problem due to the innate complexity, especially for multicomponent systems whose phase diagrams have not been established. The existing methods are either limited to binary systems or not applicable to crystalline solids at all. Here we present a new multicell Monte Carlo method based on first principles where the coexisting phases are represented by parallel supercells and free concentration search is achieved via a flip move. We first show that this method works for known miscible and immiscible systems as benchmarks and then we show that it successfully predicts the phase separation of a quaternary high entropy alloy as observed in experiments. Finally, we show that it can also work for semiconductors and intermetallics with appropriate modifiers.
11:20 AM
Boosting the CALPHAD Modeling of Multi-component Systems by ab initio Calculations: Selected Case Studies: Giancarlo Trimarchi1; Qing Chen1; 1Thermo-Calc Software AB
DFT is the main data source for the CALPHAD modeling of hypothetical multicomponent systems within the Compound Energy Formalism (CEF). However, the number of DFT calculations required by the CEF increases rapidly with the number of elements and sublattices in the system. Here, we assess the Effective Bond Energy Formalism (EBEF), a recently introduced scheme to model multi-component phases that requires fewer DFT calculations than the CEF. From a critical assessment of the EBEF, we observe that in its current formulation the parameters of the model are non-unique. Here, we propose an extra condition to ensure uniqueness of the EBEF parameters and show that the formalism so revised predicts ternary phase diagrams for the Mo-Ni-Re and Co-Cr-Re systems that agree with those obtained through the CEF calculating all ternary endmembers. We envisage that this revised EBEF will become a robust tool for DFT-driven CALPHAD modeling of multicomponent phases.