Quantum many-body systems generate high-dimensional and multiscale datasets whose physically relevant information is often distributed across subtle correlations and low-variance observables. Although machine learning and dimensionality-reduction methods offer powerful tools for exploring such data, their application to scientific problems raises fundamental challenges regarding physical validity, interpretability, and reproducibility. In the absence of ground truth, visually compelling structures may be mistaken for physical phenomena, making it necessary to establish explicit criteria for trustworthy data-driven inference. This thesis develops a visual analytics pipeline for the scientific exploration of high-dimensional quantum data, using the nonergodic extended (NEE) regime as a central case study. The NEE regime has been proposed as an intermediate behavior between thermal and many-body localized (MBL) phases, yet its physical nature remains unresolved. Rather than attempting to prove the existence of a new phase directly, this work focuses on establishing the conditions under which competing physical hypotheses can be systematically discovered, visualized, compared, and interpreted. The first stage of the pipeline addresses unsupervised discovery. By relaxing the conventional assumption of mutually exclusive ETH and MBL phases, clustering methods are employed to investigate the organization of the eigenstate manifold. While K-Means identifies structures compatible with a three-regime interpretation, Fuzzy C-Means consistently resolves only two statistically independent structures while assigning fractional memberships to intermediate states. These results suggest that the NEE regime may be interpreted either as an independent organization or as a hybrid region characterized by the coexistence of localized and thermal features, motivating the need for methodologies capable of comparing alternative explanations. The second stage establishes physical congruency as a validity condition for nonlinear visualization. Using Principal Component Analysis (PCA) as a deterministic reference, a reverse-engineering strategy is developed to evaluate whether nonlinear dimensionality-reduction methods preserve phase organization and ordering. Under this criterion, t-SNE is shown to generate physically incongruent representations, whereas UMAP consistently preserves the global organization of the quantum datasets. The third stage addresses interpretability and stability. A visual annotation framework together with quantitative reproducibility metrics is introduced to distinguish persistent geometric structures from stochastic variability across independent executions. This transforms visualization from a qualitative inspection tool into a structured analytical instrument capable of supporting reproducible interpretation. The fourth stage formalizes the reproducibility conditions required for nonlinear embeddings. By analyzing the convergence of geometric substructures across multiple datasets, a criterion based on topological consistency is established. Reproducible embeddings are shown to correspond to the emergence of a single connected structure with vanishing variance in the detected substructure count, yielding a protocol for verifiable nonlinear visualization. Finally, the complete pipeline is applied to construct a reproducible visual atlas of phase organization, transition regions, and confidence structures across the perturbation range. To the best of our knowledge, this constitutes the first continuous and reproducible visualization of eigenstate organization spanning the full range of control parameters investigated. The resulting atlas provides a common geometric representation in which physics-informed and data-driven hypotheses can be compared directly, revealing regions of agreement, ambiguity, and transitional behavior that are difficult to identify in the original high-dimensional feature space. The principal contribution of this thesis is therefore methodological. By integrating unsupervised discovery, physical congruency, interpretability, and reproducibility into a unified visual analytics pipeline, the proposed framework transforms dimensionality reduction from an exploratory visualization technique into a verifiable scientific procedure. The pipeline provides reproducible representations that preserve the global physical organization of high-dimensional quantum data while simultaneously enabling controlled exploration of subtle structures, transition regions, and low-variance phenomena that are often suppressed by conventional variance-maximization approaches such as PCA. In systems where ground truth is unavailable, this combination of stable reference representations and exploratory analysis provides a rigorous basis for hypothesis generation, comparison, validation, and future investigations involving synthetic data generation and new experimental configurations.
I sistemi quantistici a molti corpi generano dataset ad alta dimensionalità e multi-scala le cui informazioni fisicamente rilevanti sono spesso distribuite attraverso correlazioni sottili e osservabili a bassa varianza. Machine learning e i metodi di riduzione della dimensionalità offrono potenti strumenti per esplorare tali dati, la loro applicazione a problemi scientifici solleva sfide fondamentali riguardanti la validità fisica, l'interpretabilità e la riproducibilità. In assenza di una referenza ideale, strutture visivamente convincenti possono essere confuse con fenomeni fisici, rendendo necessario stabilire criteri espliciti per un'inferenza affidabile basata sui dati. Questa tesi sviluppa una pipeline di visual analytics per l'esplorazione scientifica di dati quantistici ad alta dimensionalità, utilizzando il regime esteso non ergodico (NEE) come caso di studio centrale. Il regime NEE è stato proposto come un comportamento intermedio tra le fasi termica (ETH) e localizzata a molti corpi (MBL), ma la sua natura fisica rimane irrisolta. Invece di tentare di dimostrare direttamente l'esistenza di una nuova fase, questo lavoro si concentra sullo stabilire le condizioni in cui ipotesi fisiche concorrenti possono essere sistematicamente scoperte, visualizzate, confrontate e interpretate. La prima fase della pipeline affronta la scoperta non supervisionata. Rilassando l'assunzione convenzionale di fasi ETH e MBL mutuamente esclusive, vengono impiegati metodi di clustering per studiare l'organizzazione della varietà (manifold) degli autostati. Mentre K-Means identifica strutture compatibili con un'interpretazione a tre regimi, Fuzzy C-Means identifica costantemente solo due strutture statisticamente indipendenti, assegnando appartenenze frazionarie agli stati intermedi. Questi risultati suggeriscono che il regime NEE possa essere interpretato sia come un'organizzazione indipendente sia come una regione ibrida caratterizzata dalla coesistenza di caratteristiche localizzate e termiche, motivando la necessità di metodologie in grado di confrontare spiegazioni alternative. La seconda fase stabilisce la congruenza fisica come condizione di validità per la visualizzazione non lineare. Utilizzando l'Analisi delle Componenti Principali (PCA) come riferimento deterministico, viene sviluppata una strategia di reverse-engineering per valutare se i metodi di riduzione della dimensionalità non lineare preservano l'organizzazione e l'ordinamento delle fasi. Sotto questo criterio, viene dimostrato che t-SNE genera rappresentazioni fisicamente incongruenti, mentre UMAP preserva costantemente l'organizzazione globale dei dataset quantistici. La terza fase affronta l'interpretabilità e la stabilità. Viene introdotto un framework di annotazione visiva insieme a metriche quantitative di riproducibilità per distinguere le strutture geometriche persistenti dalla variabilità stocastica attraverso esecuzioni indipendenti. Questo trasforma la visualizzazione da uno strumento di ispezione qualitativa in uno strumento analitico strutturato in grado di supportare un'interpretazione riproducibile. La quarta fase formalizza le condizioni di riproducibilità richieste per gli embedding non lineari. Analizzando la convergenza delle sottostrutture geometriche attraverso molteplici dataset, viene stabilito un criterio basato sulla coerenza topologica. Viene dimostrato che gli embedding riproducibili corrispondono all'emergere di una singola struttura connessa con varianza che tende ad annullarsi nel conteggio delle sottostrutture rilevate, fornendo un protocollo per una visualizzazione non lineare verificabile. Infine, la pipeline completa viene applicata per costruire un atlante visivo riproducibile dell'organizzazione delle fasi, delle regioni di transizione e delle strutture di confidenza attraverso l'intervallo di perturbazione. Per quanto a nostra conoscenza, questo costituisce la prima visualizzazione continua e riproducibile dell'organizzazione degli autostati che copre l'intero intervallo di parametri di controllo esaminati. L'atlante risultante fornisce una rappresentazione geometrica comune in cui ipotesi informate dalla fisica e ipotesi basate sui dati possono essere confrontate direttamente, rivelando regioni di accordo, ambiguità e comportamenti di transizione che sono difficili da identificare nello spazio delle caratteristiche originale ad alta dimensionalità. Il contributo principale di questa tesi è quindi metodologico. Integrando scoperta non supervisionata, congruenza fisica, interpretabilità e riproducibilità in un'unica pipeline di visual analytics, il framework proposto trasforma la riduzione della dimensionalità da una tecnica di visualizzazione esplorativa in una procedura scientifica verificabile. La pipeline fornisce rappresentazioni riproducibili che preservano l'organizzazione fisica globale dei dati quantistici ad alta dimensionalità, consentendo contemporaneamente un'esplorazione controllata di strutture sottili, regioni di transizione e fenomeni a bassa varianza che sono spesso soppressi da approcci convenzionali di massimizzazione della varianza come la PCA. Nei sistemi in cui una referenza ideale non è disponibile, questa combinazione di rappresentazioni di riferimento stabili e analisi esplorativa fornisce una base rigorosa per la generazione, il confronto e la validazione di ipotesi, nonché per future indagini che coinvolgono la generazione di dati sintetici e nuove configurazioni sperimentali.
Cristani, Cassio Rodrigo, A Visual Analytics Pipeline for Quantum Many-Body Systems. A Framework for Reproducible Visualization, Hypothesis Validation, and Knowledge Discovery, Borgonovi, Fausto, Tessera, Daniele, Università Cattolica del Sacro Cuore SEDE DI BRESCIA:Ciclo XXXVII [https://hdl.handle.net/10807/350662]
A Visual Analytics Pipeline for Quantum Many-Body Systems. A Framework for Reproducible Visualization, Hypothesis Validation, and Knowledge Discovery
Cristani, Cassio Rodrigo
2026
Abstract
Quantum many-body systems generate high-dimensional and multiscale datasets whose physically relevant information is often distributed across subtle correlations and low-variance observables. Although machine learning and dimensionality-reduction methods offer powerful tools for exploring such data, their application to scientific problems raises fundamental challenges regarding physical validity, interpretability, and reproducibility. In the absence of ground truth, visually compelling structures may be mistaken for physical phenomena, making it necessary to establish explicit criteria for trustworthy data-driven inference. This thesis develops a visual analytics pipeline for the scientific exploration of high-dimensional quantum data, using the nonergodic extended (NEE) regime as a central case study. The NEE regime has been proposed as an intermediate behavior between thermal and many-body localized (MBL) phases, yet its physical nature remains unresolved. Rather than attempting to prove the existence of a new phase directly, this work focuses on establishing the conditions under which competing physical hypotheses can be systematically discovered, visualized, compared, and interpreted. The first stage of the pipeline addresses unsupervised discovery. By relaxing the conventional assumption of mutually exclusive ETH and MBL phases, clustering methods are employed to investigate the organization of the eigenstate manifold. While K-Means identifies structures compatible with a three-regime interpretation, Fuzzy C-Means consistently resolves only two statistically independent structures while assigning fractional memberships to intermediate states. These results suggest that the NEE regime may be interpreted either as an independent organization or as a hybrid region characterized by the coexistence of localized and thermal features, motivating the need for methodologies capable of comparing alternative explanations. The second stage establishes physical congruency as a validity condition for nonlinear visualization. Using Principal Component Analysis (PCA) as a deterministic reference, a reverse-engineering strategy is developed to evaluate whether nonlinear dimensionality-reduction methods preserve phase organization and ordering. Under this criterion, t-SNE is shown to generate physically incongruent representations, whereas UMAP consistently preserves the global organization of the quantum datasets. The third stage addresses interpretability and stability. A visual annotation framework together with quantitative reproducibility metrics is introduced to distinguish persistent geometric structures from stochastic variability across independent executions. This transforms visualization from a qualitative inspection tool into a structured analytical instrument capable of supporting reproducible interpretation. The fourth stage formalizes the reproducibility conditions required for nonlinear embeddings. By analyzing the convergence of geometric substructures across multiple datasets, a criterion based on topological consistency is established. Reproducible embeddings are shown to correspond to the emergence of a single connected structure with vanishing variance in the detected substructure count, yielding a protocol for verifiable nonlinear visualization. Finally, the complete pipeline is applied to construct a reproducible visual atlas of phase organization, transition regions, and confidence structures across the perturbation range. To the best of our knowledge, this constitutes the first continuous and reproducible visualization of eigenstate organization spanning the full range of control parameters investigated. The resulting atlas provides a common geometric representation in which physics-informed and data-driven hypotheses can be compared directly, revealing regions of agreement, ambiguity, and transitional behavior that are difficult to identify in the original high-dimensional feature space. The principal contribution of this thesis is therefore methodological. By integrating unsupervised discovery, physical congruency, interpretability, and reproducibility into a unified visual analytics pipeline, the proposed framework transforms dimensionality reduction from an exploratory visualization technique into a verifiable scientific procedure. The pipeline provides reproducible representations that preserve the global physical organization of high-dimensional quantum data while simultaneously enabling controlled exploration of subtle structures, transition regions, and low-variance phenomena that are often suppressed by conventional variance-maximization approaches such as PCA. In systems where ground truth is unavailable, this combination of stable reference representations and exploratory analysis provides a rigorous basis for hypothesis generation, comparison, validation, and future investigations involving synthetic data generation and new experimental configurations.| File | Dimensione | Formato | |
|---|---|---|---|
|
thesis_Cassio R. Cristani_VF_Cp.pdf
non disponibili
Descrizione: Tesi dottorato
Tipologia file ?:
Tesi di dottorato
Licenza:
Non specificato
Dimensione
4.06 MB
Formato
Adobe PDF
|
4.06 MB | Adobe PDF | Visualizza/Apri |
I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.



