Physics-Consistent Calibration of Surrogate Models for Multi-Mesh Pressure Field Prediction
Please login to view abstract download link
Accurate estimation of aerodynamic loads is a critical requirement for aero-structural workflows in aircraft design, yet high-fidelity CFD simulations remain prohibitively expensive for large parametric studies. Supervised machine learning provides a computationally efficient alternative by enabling rapid prediction of surface pressure-coefficient fields, which must often be transferred across discretizations with different levels of refinement. This work presents a two-step methodology for fast multi-mesh pressure-field prediction while preserving the consistency of integral quantities. The first stage addresses surrogate-model-based interpolation across operating conditions on a reference discretization, enabling efficient inference for unseen conditions within the operational envelope. The approach is model-agnostic and can be instantiated with regression architectures widely studied in the literature [1]. The second stage focuses on multi-mesh mapping. A geometry-driven surrogate is trained to represent the pressure field for a given condition as a continuous function of geometric information, enabling evaluation on a target discretization without explicit node-to-node correspondence [2]. To ensure robustness and consistency of the mapping, the predicted field is calibrated through a constrained optimization that minimally adjusts the values while preserving integral quantities. This framework produces pressure-field predictions suitable for downstream analyses, maintaining both spatial fidelity and physically consistent global loads. The complete methodology has been validated on a three-dimensional industrial aircraft configuration, demonstrating the conservation of global aerodynamic quantities with machine-level precision. The workflow and models developed in this work are implemented within the pyLOM framework [3].
