AI and ML are commoditized in domains of human behavior, social science, marketing, and other realms that lack a "first principles" foundation. For fields that have such reference principles, such as conservation laws in physics, where there are traditional modes of scientific discovery through simulation, experiment, and theory, how can AI/ML bring additional inferential power and how can the four modalities work together better in situ?

AI and ML have largely been developed in domains where high performance is not at a premium and the main accomplishments were the refinement of the techniques. However real-time applications and data sizes that overwhelm traditional storage systems and slow down traditional algorithms are now upon us, especially from the scientific domain. How can HPC tools and techniques be leveraged to advance capabilities?

AI and ML techniques have been likened by Peter Warden of Google's TensorFlow team to "banging on the side of the TV set until it works." Their success has been amazing and profitable, but little of the mathematical rigor that accompanies other modes of discovery has been developed so far. How can we bound error, quantify uncertainty, estimate sensitivities, and accelerate convergence? Also, what fundamental mathematical, statistical, operations research, and computer science methodologies can be better placed in the service of AI/ML applications?

High-performance computing, though a “science,” depends significantly upon “art” for the tuning of algorithmic and architectural parameters of the simulation, beyond the physical parameters of the simulation model, itself. When a complex application is run repeatedly for years, e.g., in weather forecasting or aircraft design, can accumulated performance data be exposed to AI and ML tools to tune the application?