Quantifying uncertainty in machine learning for nuclear binding energy

Machine learning (ML) is rapidly becoming an important technique in nuclear physics. However, applying machine learning reliably requires quantifying the uncertainty in an ML model, which can be computationally very expensive, requiring multiple runs of training. Researchers at Lawrence Livermore National Laboratory have, for the first time in nuclear physics, implemented a method to combine training data features with biases also selected from the training data features, which allow for simultaneous fitting and uncertainty quantification. When applied to nuclear binding energies, the new method not only successfully quantifies the errors, but in particular signals when the ML model is extrapolated too far from data and becomes unreliable. This heralds a new ability to extrapolate data to extreme conditions, with reduced computational resources, yet without compromising our knowledge of how trustworthy the extrapolation is.