Providing physics guides in Bayesian neural networks from input layer: case of giant dipole resonance predictions

TL;DR

This paper demonstrates how to incorporate physics knowledge into Bayesian neural networks by selecting relevant input features, improving predictions of giant dipole resonance energies in nuclear physics.

Contribution

The work introduces a method to embed physics guides into BNNs from the input layer, enhancing prediction accuracy for nuclear data without relying on physics-motivated models.

Findings

Using ground-state properties as input reduces prediction errors.

The method effectively avoids non-physical divergence in predictions.

It helps discover physics effects from complex nuclear data.

Abstract

The Bayesian neural network (BNN) has been applied to evaluate and predict the nuclear data. However, how to provide physics guides in BNN is a key but an open question. In this work, the case study on giant dipole resonance (GDR) energy is presented to illustrate the effectiveness and maneuverability of the method to provide physics guides in BNN from input layer. The Spearman's correlation coefficients are applied to assess the statistical dependence between nuclear properties in the ground state and the GDR energies. Then the optimal ground-state properties are employed as the input layer in the BNN for evaluating and predicting the GDR energies. Those selected ground-state properties actively contributes to reduce the predicted errors and avoid the risk of the non-physics divergence. This work gives a demonstration to find effects of the GDR energy by using the BNN without the physics motivated model, which may be helpful for discovering physics effects from the complex nuclear data.