About this Abstract |
Meeting |
MS&T24: Materials Science & Technology
|
Symposium
|
Uncertainty Quantification Applications in Materials and Engineering
|
Presentation Title |
Quantitative Analysis of Systematic Uncertainties in Empirical and Machine Learning Interatomic Potentials |
Author(s) |
Amit Samanta, Collin Lewin, Vincenzo Lordi |
On-Site Speaker (Planned) |
Amit Samanta |
Abstract Scope |
With the use of machine learning (ML) techniques the field of interatomic potentials has changed drastically in the past decade. This has allowed researchers to study complex multicomponent systems at a significantly reduced computational footprint than brute-force quantum simulations. However, ML models contain a huge number of trainable parameters, and, therefore, can suffer from overfitting, and complex architectures can often obscure inherent systematic biases in these potentials. We propose a method based on Bayesian Kriging to predict systematic uncertainty in a potential. We have explored different types of kernels and features used to encode atomic neighborhood information. The proposed method is used to predict uncertainties in empirical and ML potentials available in the literature.
This work was performed under the auspices of the U.S. Department of Energy by Lawrence Livermore National Laboratory (LLNL) under Contract DE-AC52-07NA27344. This work was funded by LDRD with project tracking code 23-SI-006. |