Sustainable Approaches for Robust Predictions in Model-based Engineering Applications

Sustainable Approaches for Robust Predictions in Model-based Engineering Applications Alan Correa1, Julia Kowalski1 1Methods for Model-based Development in Computational Engineering, RWTH Aachen, Germany This work was funded by the Deutsche Forschungsgemeinschaft (DFG, German Research Foundation) – 333849990/GRK2379 (IRTG: Hierarchical and Hybrid Approaches in Modern Inverse Problems) Alan Correa [email protected] Outlook Results Approach Challenges High-Dimensional Uncertainty Propagation Heterogeneous Computational Workflows Installation and management of software dependencies Integration of components and workflow reproducibility Data management and transfer mechanisms Computational model-based Monte-Carlo methods - resource and time intensive - curse of dimensionality Surrogate model-based Monte-Carlo methods - retrain if geometry changes - parametrized by few inputs First-order Second Moment Method Reverse Mode Algorithmic Differentiation Isolation Interoperation Orchestration Showcase: Forward/Inverse Uncertainty Quantification FOSM-AD Develop a 1D/2D differentiable solver for free surface flow employing the discrete adjoint method using Enzyme.jl Benchmark accuracy and resource usage of FOSM-AD with differentiable solvers for geometric uncertainty propagation Develop a suite of model-based workflows to test for the efficacy of the proposed approach and identify challenges Determine characteristics of robust model-based workflows that are tolerant to changes in computational units and data How can we efficiently propagate uncertainties in high-dimensional input data through the model-based prediction workflow? How can we efficiently manage heterogeneity in computational units and data to develop robust model-based workflows? Package Management Containerisation Data Interchange Data Exchange Automated Data Movement Scheduling & Check-pointing Monitoring & Metrics Benchmark: FOSM-AD v/s Monte-Carlo Methods Computational Unit Data Robustify Prediction Workflows Robustness A robust prediction workflow needs to: handle variability and remain effective tolerate perturbations withstand adverse conditions Model-based Predictions Design Decision Support Development Diagnostics Applications Computational Model Input Data Pre-Process Quantity of Interest Post-Process Model Input Model Output High-Fidelity Model Surrogate Model Start End Analysis Tasks Optimization Benchmarking Model Selection Sensitivity Analysis Parameter Calibration