Towards CI-enabled, optimization-driven, simulation-based ...€¦ · Earthquake Modeling for...
Transcript of Towards CI-enabled, optimization-driven, simulation-based ...€¦ · Earthquake Modeling for...
Towards CI-enabled, optimization-driven,
simulation-based decision making
Omar GhattasUltrascale Computing LabBiomedical Engineering
Civil & Environmental EngineeringCarnegie Mellon University
Initial and boundary conditions
• Cyberinfrastructure := integrated, distributedparallel supercomputers, clusters, fast networks, federated databases, middleware, software libraries and tools, algorithms, application codes
• Design := decision making • Focus on high end, high fidelity simulation• Optimization framework
From physical models to simulation-based decision making
physical model of natural physical model of natural or engineered systemor engineered system
mathematical modelmathematical model
numerical modelnumerical modelcomputer computer simulationsimulation
validationvalidation
SimulationSimulation--based decision makingbased decision makinge.g. design, control, operations, e.g. design, control, operations,
disaster response, manufacturing,disaster response, manufacturing,hazard assessment, planning, treatmenthazard assessment, planning, treatment
scalable scalable algorithms algorithms & solvers& solvers
data/observationsdata/observations
multiscalemultiscale models models
advanced geometry &advanced geometry &discretizationdiscretization schemesschemes
parameter inversionparameter inversiondata assimilationdata assimilationmodel/data error controlmodel/data error control
visualizationvisualizationdata mining/sciencedata mining/science
optimizationoptimizationstochastic modelsstochastic modelsuncertainty quantificationuncertainty quantification
verificationverificationapproximation approximation error controlerror control
CyberInfrastructure
CyberEngineering
Earthquake Modeling for Seismic Hazard Assessment Earthquake Modeling for Seismic Hazard Assessment VolkanVolkan AkcelikAkcelik , , JacoboJacobo BielakBielak, George Biros , George Biros ((UPennUPenn), Steven Day (SDSU), Omar ), Steven Day (SDSU), Omar GhattasGhattas, , LoukasLoukas KallivokasKallivokas (Texas), Harold (Texas), Harold MagistraleMagistrale(SDSU), David (SDSU), David O’HallaronO’Hallaron
Region of interest for Region of interest for 1994 Northridge 1994 Northridge earthquake simulation earthquake simulation
Adaptive grid resolves up to 1Hz freq. Adaptive grid resolves up to 1Hz freq. w/100 million grid pts; uniform grid w/100 million grid pts; uniform grid would require 2000x more pointswould require 2000x more points
SCEC geological model provides 3D soil SCEC geological model provides 3D soil properties in Greater LA Basinproperties in Greater LA Basin
Snapshot of simulated ground motion Snapshot of simulated ground motion (simulation requires 3hr on 6Tflops PSC (simulation requires 3hr on 6Tflops PSC machine, running at >80% parallel machine, running at >80% parallel effeff))
Comparison of observation with Comparison of observation with simulation (improved prediction simulation (improved prediction requires requires petaflopspetaflops capability)capability)
Inversion of surface observations Inversion of surface observations for 17 million elastic parameters for 17 million elastic parameters (right: target; left: inversion result)(right: target; left: inversion result)
MultiscaleMultiscale Blood Flow Modeling for Artificial Heart Device DesignBlood Flow Modeling for Artificial Heart Device Design
At macroscopic (device) scales:At macroscopic (device) scales:•• Development of artificial heart assist Development of artificial heart assist
device at device at UnivUniv Pitt Med Center (Pitt Med Center (AntakiAntaki))•• Numerous advantages (size, power, Numerous advantages (size, power,
reliability, nonreliability, non--invasiveness)invasiveness)•• Design challenge: overcome tendency to Design challenge: overcome tendency to
damage red blood cellsdamage red blood cells•• Need macroscopic blood flow theory that Need macroscopic blood flow theory that
accounts for blood (cell) microstructureaccounts for blood (cell) microstructure
At microscopic (cell) scales:At microscopic (cell) scales:•• Macroscopic model fails in smallMacroscopic model fails in small--lengthlength--
scale regions (blade tip, rotor bearing)scale regions (blade tip, rotor bearing)•• Need modeling at cell scales to account Need modeling at cell scales to account
for blood damagefor blood damage•• Our Our mesoscopicmesoscopic simulations resolve simulations resolve
interaction of RBCs elastic membrane interaction of RBCs elastic membrane with plasma fluid dynamicswith plasma fluid dynamics
•• Prospects for 3D simulation of bladeProspects for 3D simulation of blade--tip tip region: 1 week at sustained 1 region: 1 week at sustained 1 petaflops/spetaflops/s
James James AntakiAntaki, Guy , Guy BlellochBlelloch, Omar , Omar GhattasGhattas, Marina , Marina KamenevaKameneva (Pitt), Robert (Pitt), Robert KormosKormos (Pitt), Gary Miller, (Pitt), Gary Miller, K. K. RajagopalRajagopal (Texas A&M), George (Texas A&M), George TurkiyyahTurkiyyah(Washington), Noel (Washington), Noel WalkingtonWalkington
Real time optimization for dynamic inversion & controlReal time optimization for dynamic inversion & control
Inversion and control for airborne contaminant transportInversion and control for airborne contaminant transport
Water network contaminant inversionWater network contaminant inversion•• Nonlinear optimization problem with Nonlinear optimization problem with
>300K variables and >100k controls >300K variables and >100k controls •• Solution time < 2 CPU minutes Solution time < 2 CPU minutes
real time source detection real time source detection •• Algorithm successful on thousands of Algorithm successful on thousands of
numerical tests on several municipal numerical tests on several municipal water networkswater networks
•• Formulation tool links to existing Formulation tool links to existing modeling software (EPANET) and modeling software (EPANET) and powerful NLP solver (IPOPT)powerful NLP solver (IPOPT)
VolkanVolkan AkcelikAkcelik, Roscoe Bartlett (, Roscoe Bartlett (SandiaSandia), Lorenz ), Lorenz BieglerBiegler, , George Biros (George Biros (UPennUPenn), Frank ), Frank FendellFendell (TRW), Omar (TRW), Omar GhattasGhattas, , Matthias Matthias HeinkenshlossHeinkenshloss (Rice), Judy Hill (CMU), David (Rice), Judy Hill (CMU), David Keyes (Columbia), John Keyes (Columbia), John ShadidShadid ((SandiaSandia), Bart van ), Bart van BloemenBloemenWaandersWaanders ((SandiaSandia), Andreas ), Andreas WachterWachter (IBM), David Young (IBM), David Young (Boeing)(Boeing)
•• sensor data provides sensor data provides concentrations of concentrations of hazardous agents hazardous agents
•• inverse problem solved inverse problem solved to reconstruct initial to reconstruct initial conditionsconditions
•• control problem solved control problem solved to find optimal to find optimal remediation strategyremediation strategy
Shape optimization of accelerator structures Volkan Akcelik (CMU) , Lori Freitag (LLNL), Omar Ghattas (CMU), David Keyes (Columbia), Patrick Knupp (SNL), Kwok Ko (SLAC), Lie-Quan (Rich) Lee (SLAC), Mark Shepherd (RPI), Tim Tautges(SNL)
CAD Meshing Partitioning(parallel)
h-Refinementp-refinement
Solvers(parallel)
Refinement
Basic Analysis Loop for given GeometryBasic Analysis Loop for given Geometry
Omega3POmega3P
S3PS3P
T3PT3P
Tau3PTau3P
•• Computer modeling has replaced trial and Computer modeling has replaced trial and error prototypingerror prototyping•• Next generation accelerators have complex Next generation accelerators have complex cavities that require shape optimization for cavities that require shape optimization for improved performance and reduced costimproved performance and reduced cost
•• Shape optimization problem governed by Shape optimization problem governed by electromagnetic electromagnetic eigenvalueeigenvalue problemproblem•• Cost functions involve target frequency, Cost functions involve target frequency, surface integrals of magnetic field, line surface integrals of magnetic field, line integrals of electric field integrals of electric field
medical medical imagingimaging
4D image 4D image registrationregistration
5D model 5D model inversioninversion
diagnosis & diagnosis & planningplanning
imaging imaging lab serverlab server
institutional institutional clustercluster
regional regional supercomputing supercomputing centercenter
physician physician desktopdesktop
ImageImage--based patientbased patient--specific inversionspecific inversion--based cardiac modelingbased cardiac modeling
VolkanVolkan AkcelikAkcelik (CMU) , George Biros (Penn), (CMU) , George Biros (Penn), AlfioAlfio BorziBorzi (Graz), (Graz), ChristosChristos DavatzikosDavatzikos (Penn), Omar (Penn), Omar GhattasGhattas (CMU), William (CMU), William GroppGropp (Argonne), Michael (Argonne), Michael HintermuellerHintermueller (Graz), (Graz), EldadEldad HaberHaber(Emory), David Keyes (Columbia), Jan (Emory), David Keyes (Columbia), Jan ModersitzkiModersitzki ((LubekLubek), ), Jennifer Jennifer SchopfSchopf (Argonne)(Argonne)
Detailed example #1: Real-time inversion for transport of contaminant
• Given local velocity field v (e.g. from mesoscopic weather model), diffusivity k, sensor observations u*, and a terrain model, estimate initial condition u0 of a convection-diffusion equation by solving regularized inverse problem:
• Forward problem can then be solved to predict transport of contaminant
Optimality system
Symbolic optimality system
Greater Los Angeles BasinOnshore flow
Transport over synthetic terrain: forward solution
• Initial concentration is a Gaussian
• Velocity field from laminar Navier-Stokes flow calculation with parabolic inflow
• Images show contours of plume at initial and after 30, 60, and 90 timesteps
• PETSc implementation
Transport over synthetic terrain: inverse solution
• Solution of inverse problem via a reduced Newton-CG solver, preconditioned by a 2-step stationary method
• Original initial condition vs. inverted initial condition for 6x6x6, 11x11x11, and 21x21x21 sensor arrays
• Solution obtained in 58 CG iterations, each requiring pair of forward/adjoint transport solves
Detailed example #2: Earthquake inversion
USGS
Variable-slip kinematicsource model
Earthquake Modeling for Seismic Hazard Assessment Earthquake Modeling for Seismic Hazard Assessment VolkanVolkan AkcelikAkcelik , , JacoboJacobo BielakBielak, George Biros , George Biros ((UPennUPenn), Steven Day (SDSU), Omar ), Steven Day (SDSU), Omar GhattasGhattas, , LoukasLoukas KallivokasKallivokas (Texas), Harold (Texas), Harold MagistraleMagistrale(SDSU), David (SDSU), David O’HallaronO’Hallaron
Region of interest for Region of interest for 1994 Northridge 1994 Northridge earthquake simulation earthquake simulation
Adaptive grid resolves up to 1Hz freq. Adaptive grid resolves up to 1Hz freq. w/100 million grid pts; uniform grid w/100 million grid pts; uniform grid would require 2000x more pointswould require 2000x more points
SCEC geological model provides 3D soil SCEC geological model provides 3D soil properties in Greater LA Basinproperties in Greater LA Basin
Snapshot of simulated ground motion Snapshot of simulated ground motion (simulation requires 3hr on 6Tflops PSC (simulation requires 3hr on 6Tflops PSC machine, running at >80% parallel machine, running at >80% parallel effeff))
Comparison of observation with Comparison of observation with simulation (improved prediction simulation (improved prediction requires requires petaflopspetaflops capability)capability)
Inversion of surface observations Inversion of surface observations for 17 million elastic parameters for 17 million elastic parameters (right: target; left: inversion result)(right: target; left: inversion result)
Inverse problem: Use records of past seismic events to improve velocity model
SCEC Phase III strong motion database: SCEC Phase III strong motion database:
Observations from 28 earthquakes and 281 stationsObservations from 28 earthquakes and 281 stations
Least squares parameter estimation formulation of inverse wave propagation
receiverssourcesdata misfit
inversion fields
displacements
forward wave propagationmodel
Multiscale inversion:Target vs. inverted isosurfaces, level 1
Multiscale inversion:Target vs. inverted isosurfaces, level 2
Multiscale inversion:Target vs. inverted isosurfaces, level 3
Multiscale inversion:Target vs. inverted isosurfaces, level 4
Multiscale inversion:Target vs. inverted isosurfaces, level 5
Multiscale inversion:Target vs. inverted isosurfaces, level 6
Multiscale inversion:Target vs. inverted isosurfaces, level 7
Multiscale inversion:Target vs. inverted isosurfaces, level 8
Multiscale inversion:Target vs. inverted isosurfaces, level 9
Comparison of target and inverted material models: 3D acoustic and elastic
Acoustic medium, pAcoustic medium, p--wave velocitywave velocity Elastic medium, sElastic medium, s--wave velocitywave velocity
Challenges/Needs
• optimization techniques for multiscalesimulation models
• optimization-ready reduced order models• large scale 4D data assimilation methods• real-time optimization algorithms • uncertainty quantification and propagation• simulation-based optimization algorithms
scalable to O(100,000) processors• latency tolerant algorithms for exploiting
distributed computing resources
SummarySummary
•• CI will help catalyze a transformation to highCI will help catalyze a transformation to high--fidelity simulationfidelity simulation--based decisionbased decision--makingmaking
•• But hardware/middleware infrastructure alone is But hardware/middleware infrastructure alone is insufficient to achieve this goalinsufficient to achieve this goal
•• Simultaneous advances on the models, methods, Simultaneous advances on the models, methods, and algorithms that underpin the components and algorithms that underpin the components ––and on their and on their systematic integrationsystematic integration to target to target strategic applications strategic applications –– are crucial for realizing are crucial for realizing the potential of CIthe potential of CI
Final editorial comment
• Advances in models and algorithms have often to led to greater improvements in simulation capability than have improvements in hardware
Example frommagnetohydrodynamics:2.5 orders of magnitude from hardware improvements; 3.5 orders of magnitude from modeling and algorithmic advances (From SCaLeSreport, Vol 2)