The design of high-performance Interior Permanent Magnet (IPM) motors with V-shape rotor topology presents a fundamental methodological challenge: classical analytical models (Magnetic Equivalent Circuits, MEC/LUMP) deliver near-instantaneous computation but break down under deep magnetic saturation, while Finite Element Analysis (FEA) provides ground-truth accuracy at prohibitive computational cost. State-of-the-art hybrid AI architectures (Zhang et al. 2025, Wang & Liu 2026) partially bridge this gap but remain constrained by a 'data-hungry' paradigm, requiring thousands of FEA snapshots for statistical training and lacking the ability to reliably handle the strongly non-linear saturation phenomena—the so-called choking effect—that occur in the thin structural bridges (ribs) of V-shape IPM rotors. This thesis proposes, implements, and validates the LUMP-PIML-GA (LPG) framework, a novel three-layer synergistic hybrid architecture that fundamentally reframes the electromagnetic synthesis problem for this class of machines. The framework integrates a physics-based Lumped Parameter Model (LUMP) as analytical backbone, a Physics-Informed Machine Learning (PIML) residual corrector, and a Genetic Algorithm (GA) inverse optimizer into a unified, self-calibrating design automation tool. This architecture is rooted in a rigorous preliminary synthesis (Chapter 5), where complex electromagnetic phenomena—including localized rib saturation, air-gap harmonic distortion, and cross-coupling effects—are systematically extracted and mapped to generate the physics-informed training foundation for the hybrid system. Novelty and Key Contributions The LPG framework advances the state of the art along four critical axes: • Data-Efficiency: The framework converges using only 12 strategically selected FEA training samples—generated via a full-factorial Design of Experiments (DoE) spanning air-gap values of 0.8–1.5 mm and current densities up to 12 A/mm²—compared to the 50+ samples required by the most data-efficient published approaches (Zhang 2025) and the hundreds to thousands demanded by standard surrogate methods. • Physical Reliability (No Black-Box): The PIML layer does not perform blind statistical regression. Constrained by Maxwell's equations, it operates exclusively on the non-linear residuals between the FEA ground truth and the LUMP prediction. This ensures that all model outputs remain strictly consistent with electromagnetic physics even at deep saturation levels (Brib > 2.0 T), where classical MEC models structurally fail. • Accuracy: The cascaded coupling of the PIML corrector with the GA optimizer reduces relative errors on quadrature flux (ψq), quadrature inductance (Lq), and radial forces/UMP (Fr) strictly below 1.75%, whereas the uncalibrated LUMP-PIML baseline prior to GA optimization exhibited a total objective function residual of 25.27%. Furthermore, the electromagnetic torque estimation error is restricted to an exceptional 0.04–0.09%. The framework's reliability is validated through a rigorous comparative benchmark against the full-fidelity FEA dataset, ensuring that torque production, saliency ratios, and radial force distributions (UMP) match the numerical ground truth across the entire operational envelope, thereby limiting torque ripple to 6.13% under peak conditions and minimizing UMP below 2 N for advanced acoustic and vibrational mitigation. • Inverse Design Paradigm: Unlike state-of-the-art models that solve the direct problem (given geometry, predict performance), the LPG framework inverts the design logic. Driven by the GA layer, it solves the inverse problem—identifying the rotor geometry and instantaneous saturation coefficients that satisfy prescribed NVH performance constraints—in the order of few seconds, enabling Digital Twin inference with an estimated 99.99% computational time saving versus transient FEA. The framework was validated against a broad set of operating conditions, of which three representative scenarios spanning the full parametric range of the study motor are reported here. Results confirm robust convergence, with relative errors consistently maintained below 2% across all key electromagnetic quantities, establishing the LPG framework as a robust and scalable Design Automation Tool for high-density IPM V-shape motors in electric vehicle traction applications.
Il progetto di motori Interior Permanent Magnet (IPM) ad alte prestazioni con topologia del rotore a V presenta una sfida metodologica fondamentale: i modelli analitici classici (Magnetic Equivalent Circuits, MEC/LUMP) offrono un calcolo quasi istantaneo, ma falliscono in condizioni di saturazione magnetica profonda, mentre la Finite Element Analysis (FEA) garantisce un'accuratezza di riferimento a costi computazionali proibitivi. Le attuali architetture ibride basate su IA (Zhang et al. 2025, Wang & Liu 2026) colmano parzialmente questo divario, ma rimangono vincolate da un paradigma "affamato di dati", richiedendo migliaia di simulazioni FEA per l'addestramento statistico e mancando della capacità di gestire in modo affidabile i fenomeni di saturazione fortemente non lineari — il cosiddetto effetto di strozzamento (choking effect) — che si verificano nei sottili ponticelli strutturali (ribs) dei rotori IPM a V. Questa tesi propone, implementa e convalida il framework LUMP-PIML-GA (LPG), un'innovativa architettura ibrida sinergica a tre strati che ridefinisce radicalmente il problema della sintesi elettromagnetica per questa classe di macchine. Il framework integra un modello a parametri concentrati (LUMP) basato sulla fisica come struttura portante analitica, un correttore dei residui basato su Physics-Informed Machine Learning (PIML) e un ottimizzatore inverso basato su Algoritmo Genetico (GA) in un unico strumento di automazione del design auto-calibrante. Questa architettura si fonda su una rigorosa sintesi preliminare (Capitolo 5), in cui complessi fenomeni elettromagnetici — inclusi la saturazione localizzata dei ribs, la distorsione armonica nel traferro e gli effetti di accoppiamento incrociato (cross-coupling) — vengono estratti sistematicamente e mappati per generare la base di addestramento physics-informed per il sistema ibrido. Originalità e Contributi Principali Il framework LPG fa avanzare lo stato dell'arte lungo quattro assi critici: • Efficienza dei dati: Il framework converge utilizzando solo 12 campioni di addestramento FEA strategicamente selezionati — generati tramite un piano di esperimenti (DoE) fattoriale completo che spazia tra valori di traferro da 0.8 a 1.5 mm e densità di corrente fino a 12 A/mm² — rispetto ai 50+ campioni richiesti dagli approcci pubblicati più efficienti (Zhang 2025) e alle centinaia o migliaia richiesti dai metodi surrogati standard. • Affidabilità fisica (No Black-Box): Lo strato PIML non esegue una regressione statistica cieca. Vincolato dalle equazioni di Maxwell, opera esclusivamente sui residui non lineari tra la "ground truth" (verità sperimentale) FEA e la predizione LUMP. Ciò garantisce che tutti gli output del modello rimangano rigorosamente coerenti con la fisica elettromagnetica anche a livelli di saturazione profonda (Brib > 2.0 T), dove i modelli MEC classici falliscono strutturalmente. • Accuratezza: L'accoppiamento in cascata del correttore PIML con l'ottimizzatore GA (Genetic Algorithm) riduce gli errori relativi sul flusso di quadratura (ψq), sull'induttanza di quadratura (Lq) e sulle forze radiali/UMP (Fr) stabilmente al di sotto dell'1.75%, mentre la baseline LUMP-PIML non calibrata, prima dell'ottimizzazione GA, presentava un residuo totale della funzione obiettivo pari a 25.27%. Inoltre, l'errore di stima della coppia elettromagnetica è limitato a un valore eccezionale compreso tra lo 0.04% e lo 0.09%. L'affidabilità del framework è validata attraverso un rigoroso benchmark comparativo rispetto al dataset FEA (Finite Element Analysis) a piena fedeltà, garantendo che la produzione di coppia, i rapporti di salienza e le distribuzioni delle forze radiali (UMP) corrispondano alla verità numerica ("ground truth") su tutto il campo operativo; ciò limita il ripple di coppia al 6.13% in condizioni di picco e minimizza l'UMP al di sotto dei 2 N per un'avanzata mitigazione acustica e vibrazionale • Paradigma di progettazione inversa e velocità in tempo reale: A differenza dei modelli allo stato dell'arte che risolvono il problema diretto (data la geometria, prevedono le prestazioni), il framework LPG inverte la logica di progettazione. Guidato dallo strato GA, risolve il problema inverso — identificando la geometria del rotore e i coefficienti di saturazione istantanei che soddisfano i vincoli prestazionali NVH prescritti — nell'ordine di alcuni secondi, consentendo un'inferenza del Digital Twin in tempo reale con un risparmio di tempo computazionale stimato del 99.99% rispetto alla FEA transitoria. Il framework è stato convalidato rispetto a un'ampia serie di condizioni operative, di cui vengono qui riportati tre scenari rappresentativi che coprono l'intero intervallo parametrico del motore in esame. I risultati confermano una convergenza robusta, con errori relativi costantemente mantenuti al di sotto del 2% per tutte le principali grandezze elettromagnetiche, stabilendo il framework LPG come uno strumento di automazione della progettazione (Design Automation Tool) solido e scalabile per motori IPM V-shape ad alta densità in applicazioni di trazione per veicoli elettrici
A novel data-efficient and physics-informed hybrid framework for high-precision electromagnetic synthesis applied to V-Shape IPM motors
Passariello, Michele
2025/2026
Abstract
The design of high-performance Interior Permanent Magnet (IPM) motors with V-shape rotor topology presents a fundamental methodological challenge: classical analytical models (Magnetic Equivalent Circuits, MEC/LUMP) deliver near-instantaneous computation but break down under deep magnetic saturation, while Finite Element Analysis (FEA) provides ground-truth accuracy at prohibitive computational cost. State-of-the-art hybrid AI architectures (Zhang et al. 2025, Wang & Liu 2026) partially bridge this gap but remain constrained by a 'data-hungry' paradigm, requiring thousands of FEA snapshots for statistical training and lacking the ability to reliably handle the strongly non-linear saturation phenomena—the so-called choking effect—that occur in the thin structural bridges (ribs) of V-shape IPM rotors. This thesis proposes, implements, and validates the LUMP-PIML-GA (LPG) framework, a novel three-layer synergistic hybrid architecture that fundamentally reframes the electromagnetic synthesis problem for this class of machines. The framework integrates a physics-based Lumped Parameter Model (LUMP) as analytical backbone, a Physics-Informed Machine Learning (PIML) residual corrector, and a Genetic Algorithm (GA) inverse optimizer into a unified, self-calibrating design automation tool. This architecture is rooted in a rigorous preliminary synthesis (Chapter 5), where complex electromagnetic phenomena—including localized rib saturation, air-gap harmonic distortion, and cross-coupling effects—are systematically extracted and mapped to generate the physics-informed training foundation for the hybrid system. Novelty and Key Contributions The LPG framework advances the state of the art along four critical axes: • Data-Efficiency: The framework converges using only 12 strategically selected FEA training samples—generated via a full-factorial Design of Experiments (DoE) spanning air-gap values of 0.8–1.5 mm and current densities up to 12 A/mm²—compared to the 50+ samples required by the most data-efficient published approaches (Zhang 2025) and the hundreds to thousands demanded by standard surrogate methods. • Physical Reliability (No Black-Box): The PIML layer does not perform blind statistical regression. Constrained by Maxwell's equations, it operates exclusively on the non-linear residuals between the FEA ground truth and the LUMP prediction. This ensures that all model outputs remain strictly consistent with electromagnetic physics even at deep saturation levels (Brib > 2.0 T), where classical MEC models structurally fail. • Accuracy: The cascaded coupling of the PIML corrector with the GA optimizer reduces relative errors on quadrature flux (ψq), quadrature inductance (Lq), and radial forces/UMP (Fr) strictly below 1.75%, whereas the uncalibrated LUMP-PIML baseline prior to GA optimization exhibited a total objective function residual of 25.27%. Furthermore, the electromagnetic torque estimation error is restricted to an exceptional 0.04–0.09%. The framework's reliability is validated through a rigorous comparative benchmark against the full-fidelity FEA dataset, ensuring that torque production, saliency ratios, and radial force distributions (UMP) match the numerical ground truth across the entire operational envelope, thereby limiting torque ripple to 6.13% under peak conditions and minimizing UMP below 2 N for advanced acoustic and vibrational mitigation. • Inverse Design Paradigm: Unlike state-of-the-art models that solve the direct problem (given geometry, predict performance), the LPG framework inverts the design logic. Driven by the GA layer, it solves the inverse problem—identifying the rotor geometry and instantaneous saturation coefficients that satisfy prescribed NVH performance constraints—in the order of few seconds, enabling Digital Twin inference with an estimated 99.99% computational time saving versus transient FEA. The framework was validated against a broad set of operating conditions, of which three representative scenarios spanning the full parametric range of the study motor are reported here. Results confirm robust convergence, with relative errors consistently maintained below 2% across all key electromagnetic quantities, establishing the LPG framework as a robust and scalable Design Automation Tool for high-density IPM V-shape motors in electric vehicle traction applications.| File | Dimensione | Formato | |
|---|---|---|---|
|
2026_07_Passariello_Michele_Tesi.pdf
solo utenti autorizzati a partire dal 25/06/2029
Descrizione: Tesi
Dimensione
9.22 MB
Formato
Adobe PDF
|
9.22 MB | Adobe PDF | Visualizza/Apri |
|
2026_07_Passariello_Michele_Executive_Summary.pdf
solo utenti autorizzati a partire dal 25/06/2029
Descrizione: Executive Summary
Dimensione
1.7 MB
Formato
Adobe PDF
|
1.7 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/10589/261495