PSimPy: GP emulation-based sensitivity analysis, uncertainty quantification, and calibration of landslide simulators

PSimPy: GP emulation-based sensitivity analysis, uncertainty quantification, and calibration of landslide simulators Hu Zhao, Anil Yildiz ([email protected]hen.de), Nazanin Bagherinejad & Julia Kowalski Methods for Model-based Development in Computational Engineering | RWTH Aachen University | ICASP14 | MS 19: Surrogate and Reduced Order Models | 12.07.2023 | Dublin, Ireland | Anil Yildiz | PSimPy: GP emulation-based sensitivity analysis, uncertainty quantification and calibration of landslide simulators | [email protected] | www.mbd.rwth-aachen.de | ICASP14 | MS 19: Surrogate and Reduced Order Models | 12.07.2023 | Slide 2 / 16 Model-based decision support © Keystone / Michael Buholzer Reference https://www.srf.ch/news/schweiz/schuttstrom-in-brienz-gr-so-sieht-es-in-brienz-gr-nach-dem-schuttstrom-aus Anil Yildiz | PSimPy: GP emulation-based sensitivity analysis, uncertainty quantification and calibration of landslide simulators | [email protected] | www.mbd.rwth-aachen.de | ICASP14 | MS 19: Surrogate and Reduced Order Models | 12.07.2023 | Slide 3 / 16 Model-based decision support Model development cycle Physical process Observation Mathematical model Computational model Verification Validation Application Model outputs Calibration Modified from Kowalski, J. 2023. Bayesian active learning for rapid flow-like geohazards. SIAM Geoscience 2023. Anil Yildiz | PSimPy: GP emulation-based sensitivity analysis, uncertainty quantification and calibration of landslide simulators | [email protected] | www.mbd.rwth-aachen.de | ICASP14 | MS 19: Surrogate and Reduced Order Models | 12.07.2023 | Slide 4 / 16 Model-based decision support Model development cycle Physical process Observation Mathematical model Computational model Verification Validation Application Model outputs Calibration Modified from Kowalski, J. 2023. Bayesian active learning for rapid flow-like geohazards. SIAM Geoscience 2023. Risk management cycle Event Prevention Mitigation Disaster preparedness Emergency response Recovery Re-analysis Anil Yildiz | PSimPy: GP emulation-based sensitivity analysis, uncertainty quantification and calibration of landslide simulators | [email protected] | www.mbd.rwth-aachen.de | ICASP14 | MS 19: Surrogate and Reduced Order Models | 12.07.2023 | Slide 5 / 16 Model-based decision support Model development cycle Physical process Observation Mathematical model Computational model Verification Validation Application Model outputs Calibration Modified from Kowalski, J. 2023. Bayesian active learning for rapid flow-like geohazards. SIAM Geoscience 2023. Risk management cycle Event Prevention Mitigation Disaster preparedness Emergency response Recovery Re-analysis Process analysis Hazard mapping Early warning Anil Yildiz | PSimPy: GP emulation-based sensitivity analysis, uncertainty quantification and calibration of landslide simulators | [email protected] | www.mbd.rwth-aachen.de | ICASP14 | MS 19: Surrogate and Reduced Order Models | 12.07.2023 | Slide 6 / 16 Physics-based (informed) ML Idea: Physics reflected in data Source: Physics-informed ML, Nature Reviews, 2021 Idea: Physics prescribed via architecture ▪Many data are needed, yet nonintrusive ▪Hard to scale to higher dimensions ▪Practical accuracy control feasible Gaussian process emulation Physics-informed neural networks Idea: Guide to physics via loss function ▪Promising, scalable results ▪Challenging to optimize hyperparameters ▪Practical error control difficult ▪GPyTorch, SciANN, DeepXDE ▪Physics exactly reflected ▪Challenging to implement ▪Feasible for moderate fidelity only CNNs, GNNs Observational bias Learning bias Inductive bias Anil Yildiz | PSimPy: GP emulation-based sensitivity analysis, uncertainty quantification and calibration of landslide simulators | [email protected] | www.mbd.rwth-aachen.de | ICASP14 | MS 19: Surrogate and Reduced Order Models | 12.07.2023 | Slide 7 / 16 Physics-based (informed) ML Idea: Physics reflected in data Source: Physics-informed ML, Nature Reviews, 2021 Idea: Physics prescribed via architecture ▪Many data are needed, yet nonintrusive ▪Hard to scale to higher dimensions ▪Practical accuracy control feasible Gaussian process emulation Idea: Guide to physics via loss function ▪Promising, scalable results ▪Challenging to optimize hyperparameters ▪Practical error control difficult ▪GPyTorch, SciANN, DeepXDE ▪Physics exactly reflected ▪Challenging to implement ▪Feasible for moderate fidelity only CNNs, GNNs Physics-informed neural networks Observational bias Learning bias Inductive bias Anil Yildiz | PSimPy: GP emulation-based sensitivity analysis, uncertainty quantification and calibration of landslide simulators | [email protected] | www.mbd.rwth-aachen.de | ICASP14 | MS 19: Surrogate and Reduced Order Models | 12.07.2023 | Slide 8 / 16 •Workflow Workflow Anil Yildiz | PSimPy: GP emulation-based sensitivity analysis, uncertainty quantification and calibration of landslide simulators | [email protected] | www.mbd.rwth-aachen.de | ICASP14 | MS 19: Surrogate and Reduced Order Models | 12.07.2023 | Slide 9 / 16 PSimPy: Predictive and probabilistic simulation with Python Sampler LHS Saltelli MetropolisHastings Simulator RunSimulator MassPointModel Ravaflow24Mixture Emulator ScalarGaSP PPGaSP Sensitivity analysis Sobol’s sensitivity analysis Inference GridEstimation MetropolisHastingsEstimation ActiveLearning •git-ce.rwth-aachen.de/mbd/psimpy •pypi.org/project/psimpy/ Anil Yildiz | PSimPy: GP emulation-based sensitivity analysis, uncertainty quantification and calibration of landslide simulators | [email protected] | www.mbd.rwth-aachen.de | ICASP14 | MS 19: Surrogate and Reduced Order Models | 12.07.2023 | Slide 16 / 16 •Reliability-managed model-based azard assessment causes bottlenecks •Computational bottleneck problems can be tackled with SciML methods •GPE is a promising method to integrate data from computational models into ML workflows •Saves significant amount of time •PSimPy can handle calibration, model selection, and sensitivity analysis efficiently •Modular structure helps to extend to other simulators Conclusions Thank you for your attention This work was partially funded by Deutsche Forschungsgemeinschaft (DFG) within the framework of the research project OptiData: Improving the Predictivity of Simulating Natural Hazards due to Mass Movements – Optimal Design and Model Selection (Project no. 441527981).