Artificial Intelligence On Nuclear Physics

Exploring Nuclear Properties Far From The Stability Line

Now days there is an urgent need for reliable, quantitative, theoretical estimates of nuclear properties far from the stability line, driven mainly by fundamental investigations of nuclear structure, as well as by experimental studies of unknown nuclei conducted at heavy-ion and radioactive ion-beam facilities, and by studies of nuclear astrophysical processes, such as s-process, r-process, or rp-process and supernova explosions^[1],[2]. Several theoretical descriptions of different levels of sophistication have been constructed over the years to determine the properties of exotic and halo nuclear systems. Although there is continuous improvement, the predictive power of these conventional, "theory-thick" models is rather limited away from the stability.

Artificial Intelligence Techniques For Statistical Modeling

The recent advance in Artificial Intelligence (AI) and especially in statistical learning theory provides, on the other hand, an alternative opportunity to develop novel, “theory-thin”, statistical approaches of finite many-body quantum systems exhibiting remarkable predictive power. Artificial Intelligence (AI) algorithms, such as Artificial Neural Networks (NNs), Support Vector Machines (SVMs) and other adaptive techniques based on statistical analysis, have the faculty to develop suitably the rich, existing collection of nuclear data^[2],[3], deconstructing the phenomenologically chaotic systematics of nuclear properties in such a manner that they correctly regenerate the known and generalize well to unknown regions, founding thus data-driven global statistical models of nuclear properties that are very useful for exploring the expanding nuclear landscape.

Research Area: Disentangling the Nuclear Puzzle

Our nuclear structure studies seek to decode the systematics of nuclear properties, focusing on generating reliable estimates of properties of exotic and halo nuclei, and nuclei of relevance to nuclear astrophysics at the limits of stability. NNs and SVMs models have been already developed by our group for a number of nuclear properties, including atomic masses, neutron separation energies, ground state spin and parities, neutron capture rates, branching probabilities in different decay channels and beta-decay halflives. It seems that our obtained models relying on AI, can match or even surpass the predictive performance of the best theory-driven approaches. Accordingly, through the gained experience of the potential of learning machines on statistical modeling of nuclear properties, this work is in progress with the prospect to construct improved models, involving more robust training algorithms and machine-learning strategies that can lead to more accurate calculations.

References

1. ^ NUPECC Long Range Plan 2004: “Perspectives for Nuclear Physics Research in Europe in the Coming Decade and Beyond ” (April 2004).
2. ^ Opportunities in Nuclear Science: “The Frontiers of Nuclear Science: A Long Range Plan” (DOE/NSF, 2007).
3. ^ G. Audi, O. Bersillon, J. Blachot, and A. H. Wapstra, Nucl. Phys. A729 (2003).