The accurate modeling of non-oscillated reactor antineutrino energy spectrum is a fundamental requirement for high precision neutrino oscillation experiments as JUNO, whose first aim is to the determine the neutrino mass ordering. The summation method, one of the primary tools for predicting such spectrum thanks to its modeling of the spectrum fine structure, is known to be affected by systematic uncertainties, culminating in the so-called 5 MeV bump. A dominant source of discrepancy is the Pandemonium effect, an experimental bias coming from incomplete nuclear decay data measured with High Purity Germanium (HPGe) gamma ray detectors, leading to mis-predictions of the spectral shape, with a systematic underestimation of the low energy part. While Total Absorption Gamma-ray Spectroscopy (TAGS) measurements provide the necessary corrections, they are available for only a limited number of fission fragments. This thesis presents a machine learning framework and methodology designed to study and correct the Pandemonium effect. A dataset was constructed, comprehensive of 52 isotopes with available TAGS measuremnts and 52 SAFE isotopes for which HPGe data is considered reliable. Various supervised regression models were trained to predict the target normalized spectra shape (TAGS when available or HPGe for SAFE isotopes) using only nuclear properties and features derived from the HPGe data. Four distinct models were developed and compared: two multitask Gaussian Process models — an Intrinsic Coregionalization Model (ICM) and a Linear Model of Coregionalization (LCM) — and two Artificial Neural Networks (a shallow and a deep one). The ICM model was identified as the best performing one, both on the single isotopes spectral shapes and on the summed spectrum of reference actinides (235U and 238U). A mitigation of Pandemonium effect is indeed obtained. This work establishes that data-driven approaches can provide robust and generalizable corrections to the summation method, and moreover highlights the impressive role they can play in the world of physics research. A valuable direction is offered by this methodology for the reduction of systematic discrepancies. The application of analogous or different data-driven techniques seems therefore an important possibility for physics, not to be overlooked in future years. It is very important however to keep in mind the intrinsic statistical structure of this framework. The predictions obtained with this methodology must be regarded as complementary tools to future experimental efforts, rather than a definitive substitute for them.
La modellazione accurata dello spettro non oscillato degli antineutrini da reattore è un requisito fondamentale per esperimenti di oscillazione dei neutrini ad alta precisione come JUNO, il cui primo obiettivo è determinare l’ordinamento delle masse dei neutrini. È noto che il summation method, uno degli strumenti principali per prevedere tale spettro grazie alla sua modellazione della struttura fine dello spettro, è influenzato da incertezze sistematiche, che culminano nel cosiddetto 5 MeV bump. Una fonte dominante di discrepanza è l’effetto Pandemonio, un bias sperimentale derivante da dati incompleti sul decadimento nucleare misurati con rivelatori di raggi gamma al germanio ad alta purezza (HPGe), che porta a previsioni errate della forma spettrale, con una sottostima sistematica della parte a bassa energia. Sebbene le misurazioni della spettroscopia di raggi gamma ad assorbimento totale (TAGS) forniscano le correzioni necessarie, sono disponibili solo per un numero limitato di frammenti di fissione. Questa tesi presenta un framework e una metodologia di apprendimento automatico progettati per studiare e correggere l’effetto Pandemonio. È stato costruito un set di dati completo di 52 isotopi con misure TAGS disponibili e 52 isotopi SAFE per i quali i dati HPGe sono considerati affidabili. Sono stati addestrati vari modelli di regressione supervisionata per prevedere la forma normalizzata degli spettri target (TAGS quando disponibile o HPGe per gli isotopi SAFE) utilizzando solo proprietà e caratteristiche nucleari derivate dai dati HPGe. Sono stati sviluppati e confrontati quattro modelli distinti: due modelli basati su processi gaussiani multitasking — un modello di coregionalizzazione intrinseca (ICM) e un modello lineare di coregionalizzazione (LCM) — e due reti neurali artificiali (una superficiale e una profonda). Il modello ICM è stato identificato come quello più performante, sia sulle forme spettrali dei singoli isotopi che sullo spettro sommato degli attinidi di riferimento (235U e 238U). Si ottiene infatti una effettiva mitigazione dell’effetto Pandemonio. Questo lavoro dimostra che gli approcci basati sui dati possono fornire correzioni robuste e generalizzabili al metodo di sommatoria e sottolinea inoltre il ruolo impressionante che possono svolgere nel mondo della ricerca fisica. Questa metodologia offre una direzione preziosa per la riduzione delle discrepanze sistematiche. L’applicazione di tecniche analoghe o meno, basate sui dati, sembra quindi un’importante possibilità per la fisica, da non trascurare negli anni futuri. È molto importante tuttavia tenere presente la struttura intrinsecamente statistica di questi modelli. Le previsioni ottenute devono essere considerate strumenti complementari ai futuri sforzi sperimentali, piuttosto che un loro sostituto definitivo.
Modeling reactor antineutrino spectrum: machine learning correction to the summation method and mitigation of the pandemonium effect
Maturi, Leonardo
2024/2025
Abstract
The accurate modeling of non-oscillated reactor antineutrino energy spectrum is a fundamental requirement for high precision neutrino oscillation experiments as JUNO, whose first aim is to the determine the neutrino mass ordering. The summation method, one of the primary tools for predicting such spectrum thanks to its modeling of the spectrum fine structure, is known to be affected by systematic uncertainties, culminating in the so-called 5 MeV bump. A dominant source of discrepancy is the Pandemonium effect, an experimental bias coming from incomplete nuclear decay data measured with High Purity Germanium (HPGe) gamma ray detectors, leading to mis-predictions of the spectral shape, with a systematic underestimation of the low energy part. While Total Absorption Gamma-ray Spectroscopy (TAGS) measurements provide the necessary corrections, they are available for only a limited number of fission fragments. This thesis presents a machine learning framework and methodology designed to study and correct the Pandemonium effect. A dataset was constructed, comprehensive of 52 isotopes with available TAGS measuremnts and 52 SAFE isotopes for which HPGe data is considered reliable. Various supervised regression models were trained to predict the target normalized spectra shape (TAGS when available or HPGe for SAFE isotopes) using only nuclear properties and features derived from the HPGe data. Four distinct models were developed and compared: two multitask Gaussian Process models — an Intrinsic Coregionalization Model (ICM) and a Linear Model of Coregionalization (LCM) — and two Artificial Neural Networks (a shallow and a deep one). The ICM model was identified as the best performing one, both on the single isotopes spectral shapes and on the summed spectrum of reference actinides (235U and 238U). A mitigation of Pandemonium effect is indeed obtained. This work establishes that data-driven approaches can provide robust and generalizable corrections to the summation method, and moreover highlights the impressive role they can play in the world of physics research. A valuable direction is offered by this methodology for the reduction of systematic discrepancies. The application of analogous or different data-driven techniques seems therefore an important possibility for physics, not to be overlooked in future years. It is very important however to keep in mind the intrinsic statistical structure of this framework. The predictions obtained with this methodology must be regarded as complementary tools to future experimental efforts, rather than a definitive substitute for them.| File | Dimensione | Formato | |
|---|---|---|---|
|
2025_12_Maturi_Executive_Summary.pdf
non accessibile
Descrizione: Executive Summary of the Thesis.
Dimensione
1.49 MB
Formato
Adobe PDF
|
1.49 MB | Adobe PDF | Visualizza/Apri |
|
2025_12_Maturi_Resoconto_in_italiano.pdf
non accessibile
Descrizione: Resoconto in lingua italiana della tesi.
Dimensione
167.98 kB
Formato
Adobe PDF
|
167.98 kB | Adobe PDF | Visualizza/Apri |
|
2025_12_Maturi_Thesis.pdf
accessibile in internet per tutti
Descrizione: Master Thesis text.
Dimensione
18.69 MB
Formato
Adobe PDF
|
18.69 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/245477