Banca de QUALIFICAÇÃO: ALLAN MOREIRA DE CARVALHO

Uma banca de QUALIFICAÇÃO de DOUTORADO foi cadastrada pelo programa.
DISCENTE : ALLAN MOREIRA DE CARVALHO
Data: 03/03/2026
HORA: 14:00
LOCAL: Auditório 801 no 8º andar do Bloco B do Campus Santo André da Universidade Federal do ABC e Webmeeting
TÍTULO:

Reduced Order Models for Parametric Domains: A Data-Driven Methodology for the Acceleration of Engineering Simulations

PÁGINAS: 152
RESUMO:

This thesis develops and validates a solver-agnostic machine-learning reduced-order
modeling (ML-ROM) framework for fast, field-level reconstruction of parametrized com-
pressible flows. While high-fidelity CFD provides accurate solutions, its computational
cost becomes prohibitive in multi-query scenarios such as design-space exploration, un-
certainty quantification, and optimization. The proposed approach address the problem
following an offline/online paradigm: high-fidelity snapshots are preprocessed and aligned,
Proper Orthogonal Decomposition (POD) is used to extract compact reduced bases, and
supervised regressors—Gaussian Process Regression (GPR) and Artificial Neural Networks
(ANNs) learn the mapping from operating and design parameters to POD coefficients,
enabling rapid reconstruction of full fields. Model selection is treated as an integral com-
ponent of the pipeline: ANN hyperparameters are optimized via Bayesian Optimization
with Hyperband (BOHB) and cross-validation, while robustness is systematically assessed
through controlled input perturbations.
A central contribution addresses geometry variability, where inconsistent meshes and
the lack of pointwise correspondence prevent direct POD application. To enable learning
on parametrized domains, a geometry-consistent alignment operator is introduced based
on a common-support representation: each geometry is mapped onto a shared parametric
domain and resampled on a fixed grid, yielding consistent snapshot vectors for POD and
enabling coherent lifting of reconstructed fields back to the physical geometry.
The framework is assessed on two complementary benchmarks. First, a supersonic
convergent-divergent nozzle problem dominated by shocks and shock-boundary-layer in-
teraction is used to compare POD-GPR and POD-ANN surrogates. A hybrid training
objective that couples reduced-space and field-space errors, together with interpretability
tools such as SHapley Additive exPlanations (SHAP) and a novel POD-mode importance
indicator, is employed to analyze model behavior. In addition, noise-robustness exper-
iments quantify performance degradation under increasing levels of perturbations and
highlight regime-dependent trade-offs: ANNs tend to preserve stability under noisy inputs,
whereas GPR can be more accurate under clean conditions but degrades faster as noise
increases. Second, POD-GPR coupled with the proposed mesh morphing operator is used
to reconstruct blade-surface pressure and temperature fields on unseen geometries of the
transonic NASA Rotor 37 compressor; the resulting ML-ROM enables stable learning
with mean test-set accuracy above R2 > 0.95 for the reconstructed fields and negligible
geometry reconstruction error.
Across both studies, the ML-ROMs reduce online evaluation from expensive CFD runs
(minutes to hours) to sub-second predictions, delivering representative speed-ups on the
order of 10^2 to 10^4 under the reported benchmark settings. The resulting methodology
provides practical guidance for building geometry-consistent, interpretable, and noise-aware
field surrogates for compressible-flow applications.

 

MEMBROS DA BANCA:
Presidente - Interno ao Programa - 2312727 - DANIEL JONAS DEZAN
Membro Titular - Examinador(a) Interno ao Programa - 1985515 - ANTONIO GARRIDO GALLEGO
Membro Titular - Examinador(a) Externo à Instituição - CAROLINA PALMA NAVEIRA COTTA - UFRJ
Membro Suplente - Examinador(a) Interno ao Programa - 1544340 - PATRICIA TEIXEIRA LEITE ASANO
Membro Suplente - Examinador(a) Externo à Instituição - LEANDRO OLIVEIRA SALVIANO - UNESP