学术交流
学术报告
报告时间： 2026年04月30日（星期四）10:00-11:00
报告地点： 四号楼 四楼中间会议室
主 持 人： 张凌
报 告 人： Darío M. Mitnik 教授
Professor, Instituto de Astronomía y Física del Espacio, CONICET-UBA, Buenos Aires, Argentina
Visiting Professor, Institute of Plasma Physics Chinese Academy of Sciences
报告简介： Large databases in atomic collision physics are growing rapidly, both from experiments and theory. But handling this massive amount of information is not trivial: simple interpolation often fails, data can be inconsistent, or may simply not exist for the specific combination we need. In this talk, we will show how machine learning can help us not only to clean and organize these data, but also to obtain new, fast, and reliable predictions. Among the examples presented, two are based on experimental compilations. One of them involves the IAEA stopping power database, the most important source of experimental stopping power data worldwide. Our group is responsible as curators of this database. We apply deep neural networks to predict electronic stopping power, after cleaning the database with an unsupervised clustering algorithm. The second experimental example is a simple yet powerful model for K-shell ionization by electron impact, which achieves excellent predictions even for elements with no experimental data. We then show two examples based entirely on theoretical calculations. On one hand, a neural network trained on continuum distorted wave (CDW) calculations for inner-shell ionization by ion impact, which replaces expensive computations with near-instant predictions. On the other hand, we apply Bayesian optimization to tune R-matrix parameters, the gold standard for electron-ion collisions, for excitation and ionization processes. This is a task traditionally done by hand and very time-consuming. The common thread of the talk is how to transform passive databases, mere compilations, into predictive knowledge engines, useful for both fundamental research and applications in plasmas, astrophysics or fusion.
