Tackling High-Dimensional Challenges in Composite Damage Modeling with Bootstrapping and Bayesian Uncertainty Quantification
Topic(s) : Special Sessions
Co-authors :
Giuseppe CATALANOTTI (ITALY)Abstract :
Composite materials exhibit inherent stochasticity, characterized by intricate damage mechanisms and numerous technological parameters that are challenging to control. The resulting randomness in their response necessitates thorough consideration during the design process. Stochastic variables, spanning geometrical and material parameters, require statistical modeling, often assumed to follow a distribution—a presumption that may prove problematic. The scarcity of experimental data further complicates the characterization of such distributions, hindering the exploration of alternative hypotheses.
The challenge deepens with error propagation, demanding resource-intensive computational models widely used in aerospace. Existing deterministic legacy models, though well-established, face hurdles in embracing stochasticity. While Monte Carlo Simulation is theoretically applicable, it becomes impractical due to computational demands. Surrogate models emerge as a solution, providing swift estimates, but their accuracy wavers with the number of retained variables, especially in high-dimensional problems.
To address these challenges, a data-driven methodology employing bootstrapping and Bayesian analysis is proposed for uncertainty quantification in high-dimensional problems with progressive damage models for composites. This approach, illustrated in a 40-dimensional case involving Open Hole Tension and Compression specimens, efficiently determines the distribution of Quantities of Interest (QoI) related to structural response. By sidestepping the need to assume distributions and surrogate models, this methodology enables the direct execution of the original computational model, enhancing its applicability in real industrial scenarios.
The challenge deepens with error propagation, demanding resource-intensive computational models widely used in aerospace. Existing deterministic legacy models, though well-established, face hurdles in embracing stochasticity. While Monte Carlo Simulation is theoretically applicable, it becomes impractical due to computational demands. Surrogate models emerge as a solution, providing swift estimates, but their accuracy wavers with the number of retained variables, especially in high-dimensional problems.
To address these challenges, a data-driven methodology employing bootstrapping and Bayesian analysis is proposed for uncertainty quantification in high-dimensional problems with progressive damage models for composites. This approach, illustrated in a 40-dimensional case involving Open Hole Tension and Compression specimens, efficiently determines the distribution of Quantities of Interest (QoI) related to structural response. By sidestepping the need to assume distributions and surrogate models, this methodology enables the direct execution of the original computational model, enhancing its applicability in real industrial scenarios.