Uncertainty and sensitivity analysis of mechanical and thermal properties computed through Embedded Atom Method potential
G Dhaliwal and PB Nair and CV Singh, COMPUTATIONAL MATERIALS SCIENCE, 166, 30-41 (2019).
Sensitivity analysis of Molecular Dynamics (MD) simulations has revealed that the predictions can be sensitive to the small perturbations in Interatomic Potential (IP) parameters. In order to make MD predictions for complex material systems more reliable, we performed uncertainty quantification of a high dimensional IP based on the Embedded Atom Method (EAM), a commonly utilized IP for metallic systems. The major contribution of this work is the prediction of a robust posterior probability distribution of the IP parameters by considering variations in the experimental values of various mechanical and thermal properties of FCC Al. The posterior probability distributions of the IP parameters were obtained using a Bayesian statistical framework. Reliability of potential parameters was assessed by performing MD simulations for a range of mechanical and thermal properties, using perturbed potential parameters. A comparison of the computed properties to existing experimental and first-principles data revealed that higher order properties such as grain boundary formation energy are sensitive (with variance of the order 10(5)) to 1% perturbations. Using a Gaussian likelihood function, a posterior probability distribution of the IP parameters that minimizes the discrepancy between MD prediction and experimental values for various mechanical properties was obtained. Final properties of interest computed using this new distribution showed less sensitivity to changes in the IP parameters. Furthermore, the obtained posterior probability distribution reflects the uncertainty due to IP parameters and the quality of MD predictions is improved by propagating that uncertainty to the final properties. Thus, instead of obtaining point valued predictions from MD, probability distributions of the final properties are obtained using this framework.
Return to Publications page