| About this Abstract |
| Meeting |
Materials Science & Technology 2020
|
| Symposium
|
Machine Learning for Discovery of Structure-Process-Property Relations in Electronic Materials
|
| Presentation Title |
Uncertainty Quantification and Active Learning of Neural Network Models for Predicting ZrO2 Crystal Energy |
| Author(s) |
Jayanth Koushik, Sungjun Choi, Aarti Singh |
| On-Site Speaker (Planned) |
Jayanth Koushik |
| Abstract Scope |
Neural networks are increasingly being used to model complex functions that arise naturally in material science; networks trained to predict crystal energies can avoid prohibitively expensive computations. However, it is challenging to analyze predictions of neural networks to guide further analysis because the prediction mechanism is poorly understood. One issue is obtaining uncertainty estimates of predictions, which can be used to identify abnormal data points, or adaptively sample additional points. We present a novel algorithm to efficiently approximate predictive variances of neural networks. Our method uses the same idea as the Jackknife method from statistics, but avoids any re-training, making it scalable to large datasets. We apply our method on a network trained to predict energy of ZrO2 crystals, and successfully identify mislabeled and abnormal structures in the data set. We also demonstrate improved performance in training the network actively, when points are sampled based on uncertainty rather than randomly. |