Application of SciML-Adapted PCMM to Deep Neural Network Surrogate Model Used for Aerodynamic Coefficient Prediction

Abstract This study applies a version of the Predictive Capability Maturity Model adapted for use with scientific machine learning (the SciML-adapted PCMM) to evaluate the credibility of a deep neural network (DNN) surrogate model used to predict aerodynamic coefficients for the NACA 0012 airfoil. The surrogate model is trained on Reynolds-averaged Navier-Stokes (RANS) simulation data across a Latin hypercube sampling of Reynolds number, Mach number, and angle of attack. Using the SciML-adapted PCMM framework, we perform a comprehensive credibility assessment across five elements: data representation, domain awareness, explainability, model validation, and uncertainty quantification. The analysis incorporates Shapley Additive exPlanations (SHAP)-based interpretability metrics, comparative validation against experimental data, and quantified numerical and surrogate model run-to-run uncertainties. Credibility levels are assigned based on rigor and maturity of each element. Areas of relative strength and credibility gaps are identified and related to the credibility level. This work demonstrates a path forward in bridging theoretical credibility models with actionable assessment protocols for SciML applications in CFD, with implications for broader adoption in safety-critical engineering domains.