Extracting Barrier Distributions from Fusion Cross Sections

Studying fusion cross sections provides insight into the fusion process, details about the internal structure of heavier nuclear systems, and a window into astrophysical processes. Extracting a “barrier distribution” from experimental cross-section data reveals details of the underlying nuclear structure. However, these distributions are numerically sensitive to the quality of experimental data available.

In this project, we build on prior work using Gaussian processes to extract barrier distributions with uncertainty estimates. We perform an extensive analysis determining how effectively Gaussian processes capture key features of the barrier distribution function as the quality of data varies. We benchmark against the analytic Wong formula for fusion cross sections. We then extend the method to explore several expressive and interpretable Bayesian Neural Network architectures in an automated fashion. We benchmark results against the Wong formula, then use our conclusions to calibrate models to experimental data and extract a wide variety of barrier distributions that can describe the data. This committee of models can deduce key characteristics of the true barrier distributions, while also identifying key regions of high uncertainty and model discrepancy to determine precisely where additional experimental data would be maximally impactful.