AI/Data Informatics: Applications and Uncertainty Quantification at Atomistics and Mesoscales: Poster Session II
Sponsored by: TMS Materials Processing and Manufacturing Division, TMS: Computational Materials Science and Engineering Committee
Program Organizers: Kamal Choudhary, National Institute of Standards and Technology; Garvit Agarwal, Argonne National Laboratory; Wei Chen, University At Buffalo; Mitchell Wood, Sandia National Laboratories; Vahid Attari, Texas A&M University; Oliver Johnson, Brigham Young University; Richard Hennig, University of Florida

Wednesday 5:30 PM
March 17, 2021
Room: RM 33
Location: TMS2021 Virtual


Multi-fidelity Machine-learning with Uncertainty Quantification and Bayesian Optimization for Materials Design: Application to Random Alloys: Julien Tranchida1; Anh Tran; Timothy Wildey1; Aidan Thompson1; 1Sandia National Laboratories
    We present a scale-bridging approach based on a multi-fidelity (MF) machine-learning (ML) framework leveraging Gaussian processes (GP) to fuse atomistic computational model predictions across multiple levels of fidelity. Through the posterior variance of the MFGP, our framework enables uncertainty quantification, providing estimates of confidence in the predictions. We used Density Functional Theory as high-fidelity prediction, while a ML interatomic potential is used as low-fidelity prediction. Practical materials design efficiency is demonstrated by reproducing the ternary composition dependence of a quantity of interest (bulk modulus) across the aluminum-niobium-titanium ternary random alloy composition space. The MFGP is then coupled to a Bayesian optimization procedure and the computational efficiency of this approach is demonstrated by performing an on-the-fly search for the global optimum of bulk modulus in the ternary composition space. This is the first application of MFGP to atomistic materials simulations fusing predictions between Density Functional Theory and classical interatomic potential calculations.

Parsimonious Neural Networks Learn Classical Mechanics and an Accurate Time Integrator: Saaketh Desai1; Alejandro Strachan1; 1Purdue University
    Machine learning tools are increasingly being utilized in the physical sciences to develop predictive data-based models as surrogates to physics-based approaches. These models have been extremely useful within restricted domains, resulting in scientific advances, but a lack of interpretability often limits their ability to extrapolate and satisfy physical invariants. We combine neural network training with genetic algorithms to find parsimonious models that describe the time evolution of a point particle under a highly non-linear potential. The genetic algorithm is designed to find the simplest, most interpretable network compatible with the training data and the resulting parsimonious neural networks discover Newton’s second law of motion expressed as a time integrator that conserves energy and is time reversible. By extracting underlying physics, the model significantly outperforms a generic feed-forward neural network and is immediately interpretable as the position Verlet algorithm, a non-trivial, symplectic integrator whose justification originates from Trotter’s theorem.

Quantifying RAMPAGE Interatomic Potentials for Metal Alloys: Elan Weiss1; Arun Hegde2; Cosmin Safta2; Habib Najm2; David Riegner1; Logan Ward1; Wolfgang Windl1; 1The Ohio State University; 2Sandia National Laboratories
    The Rapid Alloy Method for Producing Accurate General Empirical Potentials or RAMPAGE has been proposed as a computationally efficient means to generate multi-component interatomic potentials with the goal of accelerating deployment of molecular dynamics to complex alloy systems. Within this model, published EAM elemental potentials are used in conjunction with cross-interaction terms which are economically fitted to small training sets generated via DFT. RAMPAGE binary potentials can then be combined into multi-component potentials without additional fitting. By employing global sensitivity analysis, we identify uncertain parameters in RAMPAGE with dominant contributions to uncertainty in model outputs and examine their impact in different alloy systems. Using Bayesian inference, we estimate model parameters as well as model error, and compare different model constructions. We also present quantitative benchmarks of RAMPAGE potentials with respect to static equilibrium properties as well as properties of equilibrium liquids, solid solutions, and metallic glasses.

Solving Stochastic Inverse Problems for Structure-Property Linkages Using Data-Consistent Inversion: Anh Tran1; Tim Wildey1; 1Sandia National Laboratories
    In this presentation, we seek to learn a distribution of microstructure parameters that are consistent in the sense that the forward propagation of this distribution through the CPFEM model matches a target distribution on materials properties. This stochastic inversion formulation infers a distribution of acceptable/consistent microstructures, as opposed to a deterministic solution, which improves the manufacturing feasibility. To solve this problem, we utilize a crystal plasticity finite element model (CPFEM) and we employ a recently developed uncertainty quantification (UQ) framework based on push-forward probability measures, which combines techniques from measure theory and Bayes’ rule to define a unique and numerically stable solution. We combine this approach with a machine learning (ML) model based on Gaussian processes and demonstrate the proposed methodology on two representative case studies in materials design.

Use of Atomistic Based Informatics to Model Ionic Bombardment to Synthesize Boron Carbides: Kwabena Asante-Boahen1; Nirmal Baishnab2; Paul Rulis3; Michelle Paquette3; Ridwan Sakidja1; 1Missouri State University; 2University of Missouri, Columbia; 3University of Missouri, Kansas City
     In this study, we systematically modeled an important aspect of the synthesis process for a-BxC:Hy by utilizing the Reactive Molecular Dynamics (MD) in modeling the argon bombardment from the orthocarborane molecules as the precursor. The MD simulations are used to assess the dynamics associated with the free radicals that result from the ion bombardment. By applying the Data Mining/Machine Learning analysis into the datasets generated from the large reactive MD simulations, we were able to identify and quality the kinetics of these radicals. Overall, this approach allows for a better understanding of the overall mechanism at the atomistic level of Ar bombardment and the role of radical species towards the formation of the orthocarborane network and in turn the boron carbide thin films. The support from the NSF-DMREF program (Award No. 1729176) is gratefully acknowledged. Keywords: Orthocarbones, Boron carbide, Reactive MD simulations, Data Mining, Machine Learning