AI/ML in Fusion research

In this talk we will give an overview of methods that have been developed in artificial intelligence and machine learning to enhance fusion research. These often take the form of surrogate models that leverage the ability of Neural Networks (NN) to create complex representations beyond what is possible with tools from Asymptotic Analysis (a workhorse in physics and engineering). We will highlight three types of structure preserving surrogates, direct surrogates, surrogates for blended computing, and model discovery. Direct surrogates have been used to create tools for raped identification of magnetic island in Tokamaks, non-perturbative approximations to adiabatic invariants in Stellarators, and better collision operators in ICF that go beyond two body collisions, for example. In blended computing, a self-consistent mix of NN with traditional methods, they have been used for closure models for radiation transport and kinetic fluid closures that don't impose micro-scale physics. Model identification such as Weak Sparce Identification of Non-Equilibrium Dynamics has been used to identify meso-scale models from kinetic dynamics and merged with Auto Encoders to create fast surrogates for optimal design. We will end this overview with a discussion of new directions and the outlook for AI in helping curate data for next generation tools.