|About this Abstract
||2017 TMS Annual Meeting & Exhibition
||Computational Methods and Experimental Approaches for Uncertainty Quantification and Propagation, Model Validation, and Stochastic Predictions
||Automatized Convergence and Error Analyses for High Precision Density Functional Theory Calculations
||Jan Jan▀en, Tilmann Hickel, Joerg Neugebauer
|On-Site Speaker (Planned)
Over the last years methodological and computational progress in atomistic simulations has substantially improved the predictive power in materials design. However to compare the simulation results with experimental data, it is necessary to quantify the various sources of uncertainty. We therefore leverage the capabilities of our recently developed Python based workbench PyIron, to implement an automatized stochastic sensitivity analyses with the aim to determine and differentiate model errors, statistical errors and systematical errors. For each error type the convergence gradient based on our sensitivity analyses is determined and combined with the individual cost function of the parameters. Based on this function we derive an algorithm for automated convergence which allows to quantify the precision of the energy of an individual ab initio calculation as well as for derived quantities of huge sets of ab initio calculations. The efficiencey of the approach will be demonstrated for determining structural and thermodynamic quantities.