Monday Aug 22, 4pm EDT:

P. Rodriguez-Fernandez (PSFC) - Optimization projects for the advancement of fusion energy

This talk will present examples of the application of surrogate-based optimization techniques to solve fusion-relevant modeling problems, such as a gyrokinetic profile predictor and an equilibrium solver suitable for exploratory applications. The use of Gaussian Process Regression within a Bayesian optimization framework is demonstrated to accelerate the design process of fusion reactors through the development of computationally efficient modeling workflows.

N. Mandell (PSFC) - Optimizing the performance of fusion reactors with transport in the loop

As we approach the breakeven era of fusion, optimizing reactors to make them more efficient and less expensive will be critical to the wide-scale adoption of fusion as a commercial energy source. The main challenge is to achieve and maintain high steady-state pressures in the core of the reactor to reach self-sustaining fusion conditions. Since turbulence is the main source of heat transport and losses, there is an opportunity for improving reactor performance by optimizing the design for turbulent transport. In this talk, I will present a vision for tackling this challenging problem in a scalable way, consisting of three main modules: (1) fast-but-accurate turbulence modeling with GX, a GPU-native gyrokinetic code that uses pseudo-spectral (Fourier-Hermite-Laguerre) methods; (2) multi-scale modeling for predicting core profiles using a macro-scale transport solver (Trinity) coupled to many GX micro-turbulence calculations in parallel, leveraging the scale separation between turbulence and transport; and (3) transport optimization of fusion reactor designs by using (1-2) as a transport model inside the optimization loop.

Tuesday Aug 23, 4pm EDT:

A. Pavone (IPP Greifswald) - Machine learning and Bayesian modeling at Wendelstein 7-X

Bayesian inference provides an elegant framework to learn from data. Machine learning aims at developing algorithms which can learn effectively from data. I will show how the learning framework provided by Bayesian methods can support and be supported by machine learning solutions, especially modern deep learning algorithms. Through applications related to the Wendelstein 7-X fusion experiment, I will demonstrate how Bayesian inference can enhance the exploitation of interdependent heterogeneous sources of information, such as plasma diagnostic measurements, in a complex system through physics-based modeling and conventional inference methods (MCMC, MAP), and how it can benefits from recent advances based on deep learning to scale up to the large amount of data and systems found in nowadays fusion experiments. I will also introduce uncertainty quantification methods which can help making 'black-box' approaches, such as deep learning, a tool which can be relied upon in real-world applications. Here, the Bayesian framework will prove useful once again.

Wednesday Aug 24, 4pm EDT:

A. Dubey (ANL) - Software Productivity and Sustainability in Computational Science

Computational science and engineering communities develop complex applications to solve scientific and engineering challenges. These applications have many moving parts that need to interoperate with one another. These communities are facing new challenges created by the confluence of disruptive changes in computing architectures, demand for greater scientific reproducibility, and new opportunities for higher fidelity simulations with multi-physics and multi-scales. Architecture changes require new software design and implementation strategies, and significant refactoring of existing code. Reproducibility demands more rigor across the entire software endeavor. Code coupling requires aggregate team interactions including integration of software processes and practices. These challenges demand large investments in scientific software development and improved practices. In this lecture I will describe challenges of improving software productivity in computational science projects, especially those that involve high performance computing. I will also describe various efforts in mitigating these challenges through distillation of practices that have been found to be effective. The topics covered will include software design for performance portability and sustainability, and methodologies for increasing both the scientific output and the developer productivity.

Friday Aug 26, 4pm EDT:

F. Felici (SPC EPFL) - Magnetic control of TCV tokamak plasmas through Deep Reinforcement Learning

A key challenge in tokamak operations is to shape and maintain a high-temperature plasma within the vessel. This requires regulating the plasma position and shape via magnetic fields generated by a set of control coils. This work presents a new architecture for designing a tokamak magnetic controller based on deep reinforcement learning. The controller is entirely trained on a physics-based simulator and then deployed on the TCV tokamak hardware, where it was successful in controlling a diverse set of plasma configurations, including a new configuration featuring two plasmas in the vessel simultaneously. The control architecture replaces separate controllers used in traditional architectures with a single control policy. This lecture will provide details about the training and deployment of the reinforcement learning algorithm, as well as providing a comparison with more traditional control engineering solutions to the magnetic control problem.