This research project, a collaboration between data scientists and material scientists, is to design improvements in the accelerated assessment and prevention of corrosion behavior of resources and equipment through artificial intelligence and machine learning. In particular, material science studies corrosion degradation of metallic structures due to its heavy economic and maintenance burdens (~$3.2 Trillion / 4% Global GDP worldwide) with the aim to develop corrosion-resistant materials to assure their sustainability and reduce waste due to undue preventable replacement of equipment. Assessing corrosion is both time consuming and labor intensive when utilizing standard outdoor tests under natural environmental exposure conditions. Accelerated indoor corrosion tests are conducted in laboratory settings to gage performance in a shorter
period of time than outdoor tests. However, accelerated corrosion tests do not always correlate well with the actual performance in outdoor environments. We will apply deep learning and image processing methodologies to analyze associations between indoor accelerated and outdoor exposure assessments to optimize artificially accelerated methods.
This project will leverage the mobile app capability developed by WPI students for the material science engineers to record experimental data in the field, with the recorded experimental data in particular coupon images and their meta-data then uploaded and organized into a cloud data storage, providing us with rich experimental data sets. These data sets will be composed of indoor and outdoor experimental corrosion data including images, meta-data and human-generated corrosion assessments. We will apply machine learning to extract feature representations from this rich multi-media experimental data – both the human-rated indoor and outdoor corrosion along with associations between them. Second, we will develop deep models that can classify the corrosion state of an existing coupon image. Third, models for predicting the future progression behavior of corrosion for experimental data over time for distinctive coating system stacks are studied. Fourth, deep neural GAN-based networks capable of generating expected outdoor images from indoor images are to be developed. Challenges to overcome include deep learning in the context of unstructured data types, limited samples, imperfect images, human labeling bias, and missing data unpaired samples and noisy labels, for which A.I. strategies from semi-supervised to few-shot learning approaches will be explored.
Researchers: Dr. Rob Jensen, Thomas Considine, Dr. Berend C. Rinderspacher (Material Scientists, ARL), & Dr. Rundensteiner and Dr. Emdad (Data Science faculty), and Students: Biao Yin (Data science PhD), Nicholas Josselyn (Data Science PhD).
This project will leverage the mobile app capability developed by WPI students for the material science engineers to record experimental data in the field, with the recorded experimental data in particular coupon images and their meta-data then uploaded and organized into a cloud data storage, providing us with rich experimental data sets. These data sets will be composed of indoor and outdoor experimental corrosion data including images, meta-data and human-generated corrosion assessments. We will apply machine learning to extract feature representations from this rich multi-media experimental data – both the human-rated indoor and outdoor corrosion along with associations between them. Second, we will develop deep models that can classify the corrosion state of an existing coupon image. Third, models for predicting the future progression behavior of corrosion for experimental data over time for distinctive coating system stacks are studied. Fourth, deep neural GAN-based networks capable of generating expected outdoor images from indoor images are to be developed. Challenges to overcome include deep learning in the context of unstructured data types, limited samples, imperfect images, human labeling bias, and missing data unpaired samples and noisy labels, for which A.I. strategies from semi-supervised to few-shot learning approaches will be explored.
Researchers: Dr. Rob Jensen, Thomas Considine, Dr. Berend C. Rinderspacher (Material Scientists, ARL), & Dr. Rundensteiner and Dr. Emdad (Data Science faculty), and Students: Biao Yin (Data science PhD), Nicholas Josselyn (Data Science PhD).