首页|Machine-Learning Prediction of Maxwellian-Averaged Cross Sections Using Physics-Informed Nuclear Descriptors
Machine-Learning Prediction of Maxwellian-Averaged Cross Sections Using Physics-Informed Nuclear Descriptors
Hassani, Mr. Muhammad Bilal Bilal Kabir, Prof. Abdul Nabi, Prof. Jameel-Un Ali, Mr. Muhammad Jamshaid
Machine-Learning Prediction of Maxwellian-Averaged Cross Sections Using Physics-Informed Nuclear Descriptors
Machine-Learning Prediction of Maxwellian-Averaged Cross Sections Using Physics-Informed Nuclear Descriptors
摘要
Maxwellian-averaged cross sections (MACSs) are important nuclear inputs for modeling neutron-capture processes in stellar environments. Experimental MACS data are still scarce throughout the nuclear chart, however, and are especially lacking for nuclei that cannot be easily produced or measured. This work introduces a systematic machine-learning (ML) methodology for prediction of MACS based on certain physically motivated nuclear and astrophysical descriptors. A global tier-1 sample comprises 7386 nuclei, with a high experimental confidence subset of 1498 nuclei that are considered as Tier-2 in order to explore the impact of data quality on model performance. Several complementary ML methods are evaluated, including XGBoost, CatBoost, Random Forest, ExtraTrees, Deep Multilayer Perceptron, Wavelet Kolmogorov-Arnold Network (WKAN), and a physics-informed WKAN (PI-WKAN), together with a Ridge-based Super-Learner ensemble. The Super-Learner achieves optimal overall performance, with $R^2=0.9344$, $\text{RMSE} = 0.5344\text{ dex}$, and $\text{MAE} = 0.3528\text{ dex}$ for Tier-1, improving to $R^2=0.9613$, $\text{RMSE} = 0.3553\text{ dex}$, and $\text{MAE} = 0.2324\text{ dex}$ for Tier-2. Among the individual models, XGBoost provides strong performance, reaching $R^2=0.9307$ and $0.9567$ for Tier-1 and Tier-2, respectively. The feature importance and SHAP analyses indicate that proton separation energy and other types of “binding descriptors” and “structure descriptors” are the main contributors to the predictions, whereas both correlated bulk and complementary microscopic descriptors are identified by the correlation analysis. The PI-WKAN adds an explicit monotonicity constraint that is density-dependent, which gives the model a physically constrained formulation of a neural network, but with less predictive power than the best tree-based models. By integrating physically motivated nuclear descriptors with complementary ML algorithms, these results show that one can simulate available MACS data with good accuracy and offers a practical solution for estimating MACSs in experientially sparse regions in the nuclear chart.