| About this Abstract |
| Meeting |
6th World Congress on Integrated Computational Materials Engineering (ICME 2022)
|
| Symposium
|
6th World Congress on Integrated Computational Materials Engineering (ICME 2022)
|
| Presentation Title |
Machine Learning with Real-World Micrographs: A Study of Data Quality and Model Robustness |
| Author(s) |
Xiaoting Zhong, Keenan Eves, Brian Gallagher, Yong-Jin Thomas Han |
| On-Site Speaker (Planned) |
Xiaoting Zhong |
| Abstract Scope |
In this presentation we show how data quality affect machine learning (ML) models. In particular, we evaluate ML models developed to predict mechanical properties of a molecular solid, 1,3,5-triamino-2,4,6-trinitrobenzene (TATB), using SEM micrographs. Different sets of micrographs were collected with various brightness and contrast settings and stationary microstructure content. Several feature descriptors from different feature classes were applied to encode the micrographs. A random forest model was then trained to predict the ultimate compressive stress of consolidated TATB samples. Results show that instrument-induced pixel intensity signals can be encoded into micrograph feature descriptors and affect ML model predictions in a consistently negative way. Image standardization methods, like histogram equalization, were applied to reduce the instrument-induced signals and successfully improved micrograph feature quality. We conclude that either a careful control of micrograph quality or a careful choice of ML pipeline is necessary to create reliable quantitative micrograph analysis using ML approaches. |
| Proceedings Inclusion? |
Definite: Other |