Электронный каталог

Статья

Talochka, E.
Enhanced KAN Architecture for Experimental Data Processing in High-Energy Physics / E.Talochka, G.Ososkov, N.Voytishin, [a.o.]. – Text : electronic // Computer Physics Communications. – 2025. – Vol. 316. – P. 109801. – URL: https://doi.org/10.1016/j.cpc.2025.109801. – Bibliogr.: 25.

An enhanced Kolmogorov-Arnold Network (KAN) compatible with the Adam optimizer is developed and applied to the deconvolution problem of multi-Gaussian signals and the fitting problem of the 3D distribution of the magnetic field in the BM@N (Baryonic Matter at Nuclotron) spectrometer of the Nuclotron-based Ion Collider fAcilit (NICA). Stable training dynamics and rapid convergence with the Adam algorithm, closely matching those of the computationally intensive LBFGS method, are achieved by implementing activation functions as a superposition of asymmetric super-Gaussian components and initializing their weights close to zero. The proposed KANs exhibit high accuracy (> 90 %) in the deconvolution of overlapping Gaussian signals with an unknown number of components as well as in the modeling of complex magnetic field geometries.



Спец.(статьи,препринты) = С 344.1ш - Методы обработки результатов измерений
Спец.(статьи,препринты) = Ц 840 в - Программы обработки экспериментальных данных и управление физическими установками$
ОИЯИ = ОИЯИ (JINR)2025