Banca de DEFESA: ALLAN MOREIRA DE CARVALHO

Uma banca de DEFESA de DOUTORADO foi cadastrada pelo programa.
DISCENTE : ALLAN MOREIRA DE CARVALHO
Data: 03/08/2026
HORA: 09:00
LOCAL: https://meet.google.com/gjw-vczk-sqf
TÍTULO:

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

PÁGINAS: 295
RESUMO:

This thesis investigated the fast reconstruction of parametrized compressible-flow fields using machine-learning reduced-order modeling (ML-ROM). The work addressed the high cost of repeated computational fluid dynamics (CFD) evaluations in parametric studies, design exploration, and optimization. The main contribution was a solver-agnostic workflow that combines proper orthogonal decomposition (POD), Gaussian process regression (GPR), artificial neural networks (ANNs), mesh morphing, and diagnostic analyses within a reproducible offline/online methodology for field-level surrogate modeling.

Two compressible-flow case studies were considered. The first was a convergent–divergent nozzle with shock-dominated internal flow. In this case, the surrogate models were trained to approximate high-fidelity two-dimensional (2D) Reynolds-averaged Navier–Stokes (RANS) fields for pressure, temperature, and Mach number. A relevant contribution was the use of lower-fidelity quasi-one-dimensional (Q1D) Euler fields as physics-informed surrogate inputs. Unlike purely scalar operating parameters, these fields encoded information about the geometry and the flow distribution, allowing the surrogate to generalize to unseen area distributions that were not directly represented by the scalar parameterization. The analyses also showed that increasing the POD basis did not indefinitely improve the end-to-end prediction, since the surrogate performance saturated once regression error became dominant. ANN models were more effective in exploiting lower-energy POD modes for flow reconstruction, whereas GPR models provided faster and smoother predictions in favorable regimes.

The second case study was the National Aeronautics and Space Administration (NASA) Rotor 37 transonic compressor under geometry-parametric blade variations. This case addressed a key limitation in applying POD to parametrically deformed domains: the absence of pointwise correspondence among non-matching three-dimensional (3D) surface meshes. To overcome this issue, a geometry-consistent common-support representation was used to map the blade surfaces onto a shared parametric domain before POD compression and POD–GPR regression. The resulting surrogate reconstructed pressure, temperature, and surface geometry on unseen blade configurations with high field-level accuracy and reduced online evaluation time by approximately four orders of magnitude.

Across both case studies, POD-rank sensitivity, snapshot convergence, perturbation stability, robustness tests, and interpretability tools were used to diagnose the reliability of the reduced representations. The thesis therefore established an ML-ROM methodology for accelerating repeated CFD-based field evaluations in compressible-flow applications. The proposed framework was shown to act as a complementary layer to high-fidelity CFD, supporting fast parametric studies, sensitivity analyses, preliminary design exploration, and optimization-oriented workflows while retaining access to spatially resolved flow information.

MEMBROS DA BANCA:
Presidente - Interno ao Programa - 2312727 - DANIEL JONAS DEZAN
Membro Titular - Examinador(a) Interno ao Programa - 1544340 - PATRICIA TEIXEIRA LEITE ASANO
Membro Titular - Examinador(a) Externo à Instituição - FRANCO JEFFERDS DOS SANTOS SILVA - ITA
Membro Titular - Examinador(a) Externo à Instituição - ALBERTO COSTA NOGUEIRA JUNIOR - UNICAMP
Membro Titular - Examinador(a) Externo à Instituição - RODRIGO VIDONSCKY PINTO - IFSP
Membro Suplente - Examinador(a) Interno ao Programa - 1985515 - ANTONIO GARRIDO GALLEGO
Membro Suplente - Examinador(a) Externo à Instituição - LEANDRO OLIVEIRA SALVIANO - UNESP