Nuclear Physics in Astrophysics XII (NPA-XII)

Nuclear Mass Predictions for Astrophysical Applications using Ensemble Learning and Model Averaging (ELMA)}

Not scheduled

5m

Cluj-Napoca, Babeş – Bolyai University

Speaker

Ms Nisha Chandnani (Department of Physics, School of Physical and Biological Sciences, Manipal University Jaipur, Jaipur 303007, India)

Description

Accurate nuclear mass predictions are essential for astrophysical reaction network calculations, particularly for $r$-process nucleosynthesis, where uncertainties of even a few hundred keV can alter elemental abundances by orders of magnitude [1, 2]. Global theoretical mass models have progressively reduced their RMSE to the range of 0.2--0.8 MeV [3-8], yet their predictive reliability deteriorates significantly for nuclei far from the $\beta$-stability line, well above the sub-100 keV precision required for high-fidelity $r$-process calculations [9, 10].

We present ELMA (Ensemble Learning and Model Averaging), a framework that combines the Gradient Boosting Regressor (GBR) [11] with a normalized weighted averaging scheme to improve nuclear mass predictions. Six nuclear mass models spanning macroscopic-microscopic and fully microscopic frameworks: WS4, WS4+, FRDM, DZ(28), UNEDF1, and RMF [3-8] are independently corrected using GBR trained on their raw residuals with respect to the AME2020 dataset [12]. The corrected residuals are combined through a weighted average, where weights are assigned inversely proportional to each model's RMSE, ensuring that better-performing models contribute more to the final prediction. The weighted averaging leverages the partial cancellation of model-specific deviations: the corrected residuals of individual models are randomly scattered about zero with opposite signs, leading to a noticeable reduction in the net error of the ensemble prediction [13, 14].

The resulting ELMA model achieves an RMSE of $\sim$57 keV for the complete AME2020 dataset [12], well below the critical threshold of 100 keV. The GBR correction substantially redistributes the model weights, enabling all six models to contribute more comparably to the ensemble and reducing sensitivity to any single model's bias. The validity of ELMA is demonstrated through evaluation of $Q$ values for $\alpha$ decay, showing a marked reduction in deviations from experimental data. Nuclear mass excesses and binding energies for $\sim$6300 nuclei are made publicly accessible via https://ddnp.in.

References:
[1] M. Arnould et al., Phys. Rep. 450 (2007) 97.
[2] D. Martin et al., Phys. Rev. Lett. 116 (2016) 121101.
[3] P. Möller et al., At. Data Nucl. Data Tables 109 (2016) 204.
[4] N. Wang et al., Phys. Lett. B 734 (2014) 215.
[5] N. N. Ma et al., Chin. Phys. C 43 (2019) 044105.
[6] J. Duflo, A. P. Zuker, Phys. Rev. C 52 (1995) R23.
[7] M. Kortelainen et al., Phys. Rev. C 85 (2012) 024304.
[8] L. S. Geng et al., Prog. Theor. Phys. 110 (2003) 921.
[9] R. Utama, J. Piekarewicz, Phys. Rev. C 96 (2017) 044308.
[10] M. Mumpower et al., Prog. Part. Nucl. Phys. 86 (2016) 86.
[11] J. H. Friedman, Ann. Stat. 29 (2001) 1189.
[12] W. J. Huang et al., Chin. Phys. C 45 (2021) 030002.
[13] V. Kejzlar et al., Sci. Rep. 13 (2023) 19600.
[14] E. Alhassan et al., Nucl. Sci. Tech. 35 (2024) 205.