Abstract: Machine learning approaches have demonstrated state-of-the-art performances in a wide range of commercial application domains; however, their success is limited for scientific domain because of several challenges, such as uncertainty, data scarcity, and model interpretability.
In the first part of the talk, I will describe probabilistic machine learning approaches (sensitivity-based Gaussian process and a conditional variational autoencoder (CVAE) developed to model application I/O performance and its variability as a function of application and file system characteristics on leadership-class systems. Then I will describe a concept-drift-aware predictive modeling approach to adapt the performance models to account for abrupt changes in high-performance computing systems. This approach consists of online Bayesian changepoint detection and a moment-matching transformation to adapt the predictive model for the changepoint.
In second part of the talk, I will describe my work on deep learning for image data from physical sciences. First, I will describe our approach to automatically segment 3-D atom probe tomography data of cobalt- and aluminum-based superalloys, using deep-learning-based edge detection that transfer-learn knowledge from real world images. Then, I will describe deep-learning-based compression artifact reduction in JPEG images of outputs generated from climate simulations, which can transfer-learn knowledge from previous simulations to dynamically enhance the data from the running simulation.
I will conclude by touching upon other ongoing work and discuss my future research directions in scientific machine learning.