December 18, 2020
Francesco Sciortino is a PhD candidate in the MIT Department of Physics, working at the Plasma Science and Fusion Center (PSFC) under the direction of research scientist Nathan Howard and PSFC associate director Earl Marmar. He recently gave an invited talk at the High Temperature Plasma Diagnostics (HTPD) conference (December 14-17, 2020). His research focuses on plasma turbulence in tokamaks, with an emphasis on computational statistics and machine learning methods.
What research question did you explore?
High Temperature Plasma Diagnostics (HTPD) is a conference about “plasma diagnostics”, generally interpreted by most as “instrumentation.” However, the work that I presented is more aligned with “data science” than “instrumentation.” This is a line of research that the HTPD organizers have been trying to highlight at the conference in recent years. The reason is that today’s most advanced experimental research is often a combination of modern hardware and cutting-edge software, with machine learning becoming an ever-more-important part of the story. My HTPD talk presented the work that came from a collaboration between me and Norman Cao (MIT PhD ’20, now at New York University). It was initially a “side project” for both of us, potentially useful for both of our PhD research paths, but over time we expanded it and it became important for my research. It was eventually published in IEEE Transactions on Plasma Science, in a Special Issue on Machine Learning, Data Science and Artificial Intelligence in Plasma Research.
I think that the HTPD organizers offered an invited talk in recognition of the fact that our paper has some appealing generality to it: its methods and principles are quite broadly applicable. Our paper describes a modern solution to an old problem: the analysis of atomic lines that are complicated by overlaps, unclear identification and numerous other issues. These problems are particularly important for high-resolution spectroscopy, both in laboratory plasmas (e.g. in tokamaks) and in astrophysics.
How does your research address this question?
The astrophysics community has made tremendous progress on applying modern data science, and we borrowed some of their best tools and combined them with an elegant solution to the problem that we faced in our own data. We focused on experimental measurements from the X-ray Imaging Crystal Spectrometry (XICS) diagnostic at Alcator C-Mod (MIT’s fusion experiment, retired in 2016). The tools that we developed, part of the “Bayesian Spectral Fitting Code” (BSFC), attempt to separate multiple overlapping, quasi-Gaussian atomic line shapes, taking care to quantify uncertainties in detail and rigorously determine how complex the fitting process should be. We decomposed each line shape into a Hermite polynomial series, allowing minimal numbers of free parameters in the learning task. The methods that we applied are part of the expanding field of Bayesian statistics, which offers a principled approach to uncertainty, parameter estimation and model selection.
What are the next steps?
While BSFC is only trying to solve a relatively specific task, its methods are likely to find more and more applications in experimental data analysis. In my HTPD talk, I tried to provide a good motivation to why scientists should consider our methods in the context of their own research tasks.
I will soon submit a new paper to expand on applications of BSFC to multi-line spectral fitting to infer impurity transport coefficients in tokamak plasmas, which is a major subject of my PhD thesis.