Speaker
Description
The rise of deep learning has provided transformative tools across numerous scientific disciplines, including nuclear physics, and has attracted significant attention from researchers. However, the 'black box' nature of deep learning often raises concerns about its reliability, particularly in critical applications such as nuclear physics. The reliability of models comes not only from accurate predictions but also from well-calibrated uncertainty quantification and the ability to extrapolate beyond the available data. In this talk, we discuss these fundamental challenges in conventional deep learning and demonstrate how advanced techniques developed by recent computer science progress can effectively address them. We first introduce straightforward yet scalable and effective uncertainty quantification methods grounded in probabilistic frameworks, showcasing their application in R-matrix fitting to nuclear elastic scattering. Furthermore, we present a robust solution to overcome the inherent limitation of extrapolation by leveraging neural networks to uncover underlying mathematical expressions, highlighting its success in predicting nuclear properties across the landscape. Our work underscores that, with the right techniques, deep learning can serve as both a powerful and trustworthy tool in nuclear physics research.
