Search

Mellin-Barnes representation of GPDs is presented and how it can be used for fitting procedure to experimental data.

It is shown how conformal symmetry is used to calculated next-to-leading and next-to-next-to-leading corrections to amplitude of deeply virtual Compton scattering.

Duboko virtualno komptonsko raspršenje (DVCS) je proces kojim se dobro može izučavati trodimenzionalna raspodjela kvarkova i gluona u protonu. Funkcije strukture tog procesa (komptonski form faktori, CFF-ovi) su esencijalno neperturbativni objekti koje modeliramo neuronskim mrežama. One predstavljaju optimalan izbor zbog činjenice da ne unose pristranost. U ovom radu, tako generiran model DVCS amplituda prilagođava se različitim eksperimentalnim mjerenjima. Pokazano je da je prilagodba na podatke najkvalitetnija u slučaju kada se koriste svi CFF-ovi u vodećem doprinosu. Uočene su opservable na koje se model kvalitetnije prilagodi u odnosu na druge te su prikazani naučeni CFF-ovi ovisno o izboru značajki neuronske mreže. Deeply virtual Compton scattering (DVCS) is a process which provides access to quality research of three-dimensional distribution of quarks and gluons inside the proton. Structure functions of that process (Compton form factors, CFF) are essentially non-perturbative objects which we model with neural networks. They represent the optimal choice due to the fact that they do not introduce bias. In this thesis, DVCS amplitudes generated that way are fitted to different sets of experimental data. It is shown that using all leading order CFF functions in modeling experimental observables yields best fit to the data. Also, observables which yield better fits to data than the others are notices and learned CFF functions, based on neural network features choices, are shown.

U ovom radu promatrat će se mogućnost ekstrakcije komptonskih form faktora (CFF funkcija) iz tipičnih opservabla mjerenih u procesima leptoprodukcije fotona, gdje su CFF funkcije modelirane neuronskim mrežama. Kako bi se provjerilo hoće li postupak učenja CFF funkcija neuronskim mrežama biti uspješan i o čemu ovisi kvalitetna ekstrakcija, generirani su umjetni eksperimentalni podaci (opservable na koje je dodan šum) iz poznatog Goloskokov-Kroll modela za CFF funkcije. Pokazano je da se uz dovoljan broj podataka, CFF funkcije mogu naučiti, da je kvaliteta ekstrakcije to bolja što ima više podataka te su uočene opservable čije uključivanje u skup podataka za učenje poboljšava kvalitetu ekstrakcije u većoj mjeri nego uključivanje drugih opservabli. In this thesis we will study a possibility of extraction of Compton form factors (CFFs) from typical observables measured in the process of leptoproduction of photon, where the CFF functions will be modeled with neural networks. In order to verify the validity of learning CFF functions using neural networks and to determine what makes a quality extraction, artificial (mock) experimental data (observables to which noise was applied) was generated from popular Goloskokov-Kroll model for CFF functions. It is shown that with the enough data, CFF functions can be learned, that the quality of extraction is getting better when increasing the number of data and some observables were noticed which had more effect on the quality of extraction than the others.

We report on some new theoretical results for the representation of generalized parton distributions (GPDs), the understanding of next-to-leading order corrections to hard exclusive processes, the calculation of higher twist corrections to deeply virtual Compton scattering, and the progress in describing experimental measurements in terms of GPD models.