Guojing Cong

My recent work focuses on several related areas. Intelligent simulations: The aim is to accelerate scientific discovery through AI augmented simulations. We apply machine learning and data analytics to steer, guard, and accelerate simulations. Some current projects include state cataloguing and anomaly detection in simulating the Ras proteins on cell membranes for cancer research (with Lawrence-Livermore National Lab), mesh entanglement prediction in arbitrary Lagrangian-Eulerian method (With Lawrence-Livermore National Lab), molecule affinity prediction (with IBM Computational Biology Group), dynamic time-step optimization in multi-sale platelet dynamics modeling (with StonyBrook University), and surrogate models. AI in science brings unique challenges. For example, reducing the data to images or videos can result in loss of important structural information; in-situ analysis demands fast on-line and incremental training; and deployment onto modern architectures needs to coordinate complex data paths between simulation and AI. Distributed training algorithms and implementations. We aim to design efficient distributed training algorithms that do not demand frequent communications with fast convergence. One of our papers show that K-step averaging does not need frequent communications for fast convergence. Deep learning applications, either alone or embedded into other workflows, are typically big-data applications. We study I/O optimizations and system support (e.g., NVMe over RDMA) for these systems. Finally, in pushing these workloads to clouds, we also investigate system software support on the cloud infrastructure.