Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Karthik Duraisamy
May 18 2017, Univ of Michigan
ConFlux: An Ecosystem for Data-enabled Computational Physics
“Commercial” applications
Data
Predictive capability
Machine Learning
• No physical law ;• Data is directly useful for model;• Large amounts of relevant data.
Data, UQ, Physics and all that..
SEQUOIA Team
3
Data, UQ, Physics and all that..
SEQUOIA Team
4
Not really !
Data+
Inference +
Physical model+
Machine Learning+
Theoretical insight +
Problem-specific thought process
+Computer science
= Useful solution
Data &Comp. Science
High perf. computing
Physics & Modeling
Why may this not work in physical systems?
Predictive capability
Machine Learning
• We may never have enough data to address the entire range of physical phenomena
• Boundary conditions, physical constraints, etc.
Data
Combining Data and Physical Modeling
Truth
Model
Data-augmented
model
Combining Machine Learning and Physical Modeling
Predictive capability
Physical Model + Machine Learning
• We will never have enough data to address the entire range of physical phenomena Use physics model to pin data and
constrain it
• Data contains real quantities; Model contains “modeled” quantities (loss of consistency)
• Data will be only loosely connected to model (and not objective)
• Data will be noisy and of variable quality
Data
Field Inversion + Machine learning to Augment Physics-based, Consistent Models
Data
Predictive capability
Physical model + Inference
Machine learning
Physical model + consistent augmentation
• We will never have enough data to address the entire range of physical phenomena Use physics model to pin data and
constrain it
• Data contains real quantities; Model contains “modeled” quantities (loss of consistency) Inference connects real quantities to
modeled ones
• Data will be only loosely connected to model (and not objective) Inference connects secondary, non-
objective data to model quantities
• Data will be noisy and of variable quality Probabilistic casting of inference and
learning
Inverse Problem 1
Gd1
Inverse Problem 2
Gd2
Inverse Problem n
Gdn
Machine Learning / Big Data
PDF / Map of β (η)
Dataset 1 Dataset 2 Dataset n Schematic of Framework
Extrapolation capability
(from theory, existing models)
δ1, η1 δ2, η2 δn, ηn
Inverse Problem 1
Gd1
Inverse Problem 2
Gd2
Inverse Problem n
Machine Learning / Big Data
PDF / Map of δ (η)
Predictive model
M(Q,Qt,Qx, δ(η),..) =0
ηQuery
Realization
Pre-processing Prediction (for one realization)
Dataset 1 Dataset 2 Schematic of Framework
Extrapolation capability
(from theory, existing models)
Dataset n
Gdn
δ1, η1 δ2, η2 δn, ηn
δ
ConFlux: A Novel Platform for Data-driven Computational Physics
Big Data Machine LearningTraining Tier
Spark, Hadoop, GraphLab, BlinkDB, DBMS, ...
4 TB RAM
Spark, Hadoop, GraphLab, etc for scalable machinelearning training aggregation analytics on massive in-memory and out of core datasets.
Training Tier
Machine LearningTesting /HPC Tier
Pascal GPU + NVLink
Receive results from training Tier and perform massively parallel evaluations (e.g. GP)
(CAPI)/NVLink to allow low Latency data transfers between GPUs and CPUs.
Testing Tier
Big Data Machine LearningTraining Tier
Spark, Hadoop, GraphLab, BlinkDB, DBMS, ...
4TB RAM
IBM Elastic Storage System (GPFS)
HDFS / MPI-IO / Posix
SSD Buffer 1PB HDD
Two subsystems high throughput bulk storage holding I/O of largesimulations. low latency storage for
check-pointing andtransactional model updates during different iterations of the physical modeling.
IBM ESS supporting GPFS high-throughput I/O withboth traditional and HDFS compatible access.
Storage Tier
Big Data Machine LearningTraining Tier
Spark, Hadoop, GraphLab, BlinkDB, DBMS, ...
4TB RAM
Machine LearningTesting /HPC Tier
Pascal GPU + NVLink
IBM Elastic Storage System (GPFS)
HDFS / MPI-IO / Posix
SSD Buffer 3PB HDD
Big Data Machine LearningTraining Tier
Spark, Hadoop, GraphLab, BlinkDB, DBMS, ...
4TB RAM
Machine LearningTesting /HPC Tier
Pascal GPU + NVLink
Resource Manager & Scheduler
IBM Platform Symphony
Seamless, multi-tenant andlow-latency schedulingacross heterogeneous HPCand Hadoop-based clusters.
Manage both YARN-based components (e.g., Hadoopnodes, Spark nodes, etc.) aswell as HPC nodes
Scheduling Tier
IBM Elastic Storage System (GPFS)
HDFS / MPI-IO / Posix
Big Data Machine LearningTraining Tier
Machine LearningTesting /HPC Tier
Resource Manager & Scheduler
SSD Buffer 1PB HDD
Spark, Hadoop, GraphLab, BlinkDB, DBMS, ...
4TB RAMPascal GPU + NVLink
Throughput Big Data Interconnect (100Gbps speed)
IBM Platform Symphony
Control Interconnect (1Gbps)
InfiniBand interconnects to enable both low-latency and high-throughput transfers
TCP drivers to ensurecapability for ourHadoop-based components.
Interconnects
IBM Elastic Storage System (GPFS)
HDFS / MPI-IO / Posix
Big Data Machine LearningTraining Tier
Machine LearningTesting /HPC Tier
Resource Manager & Scheduler
Commodity HPC cluster (Flux, XSEDE, Globus-Transfer) SSD Buffer 1PB HDD
Spark, Hadoop, GraphLab, BlinkDB, DBMS, ...
4TB RAMPascal GPU + NVLink
Throughput Big Data Interconnect (100Gbps speed)
IBM Platform Symphony
WAN 100Gbps Internet 2 Service via
UM Science DMZ
Control Interconnect (1Gbps)
Example: Turbulence Modeling for Aerodynamics
Singh, A., Medida, S. & Duraisamy, K., Data-augmented Predictive Modeling of Turbulent Separated Flows over Airfoils, AIAA Journal, 2017
Models perform poorly in separated flows.
True prediction
Singh, A., Medida, S. & Duraisamy, K., Data-augmented Predictive Modeling of Turbulent Separated Flows over Airfoils, AIAA Journal, 2017
Prediction – S805
Collaboration with Altair, Inc.
True prediction !
Inference used only CL data, NN-augmented model provides considerable predictive improvements of Cp
S 809, Re=2 Million
Variability
α=0
α=14 α=20
Training from different sets
S 809, Re=2 Million
Portability : Implementation in AcuSolve
S809 Airfoil : Predictive results in Commercial CFD solver
Vision for the futureA continuously augmented curated database / website of inferred corrections that are input to the machine learning process
Users upload/download/process data, generate maps.
• Traditionally, models have been provided through:– A complete set of PDEs with modeled closure terms– A printed article to deliver all this content
• In the future, we expect to provide:– A set of PDEs comprising all that is well known about the
behavior of the turbulence– An auxiliary piece of software that can be embedded into a
RANS solver and that contains the machine-learned closure terms (appropriately version controlled)
– A software repository with clear explanations of the datasets that were used to create the “model”
– Possibly multiple versions of the “models” that have been trained with different datasets and are more appropriate for different flow conditions.
Turbulence models of the future
Precipitate Morphology
G. Treichert & K. Garikipati (Mech Engineering and Materials Science)
Material properties can be significantly improved through alloying and its associated precipitate formation.
● The geometry of the precipitate affects the resulting material properties, and it is driven largely by the minimization of the interfacial and strain energies in the precipitate.
● Current methods rely on phase field models, which involve solving time dependentPDEs, using inherently serial time stepping schemes.
Precipitate Morphology
G. Treichert & K. Garikipati (Mech Engineering and Materials Science)
Materials ModelingThe goal is to identify, explain, predict and ultimately to design the properties and responses of these materials.
Hierarchical models have been developed at several scales These methods have thus far provided insight and qualitative connections to
parameters and phenomena from lower scales, but have not been predictive
Quantum Monte Carlo Density Functional Theory Continuum physics
Profs. Vikram Gavini and Krishna Garikipati (Mech Engineering and Materials Science)
Subject-specific blood flow modeling
Biggest challenges lack of physiologic data to
inform the boundary conditions lack of data on mechanical
properties of the vascular model
Obtain data from tomography and MRI
Solve inverse problem for parameters
Massive data size
On-the-fly Lagrangian computation of Motion
Evaluation of arterial stiffness from medicalImages !
Prof. Alberto Figueroa (Biomedical Engineering & Surgery)
Climate system interactionsThe Earth's climate system is composed of multiple interacting components that span spatial scales of 13 orders of magnitude and temporal scales that range frommicroseconds to centuries. key responses and feedbacks in the system are not well characterized
Understanding how clouds interact with thelarger scale circulation, thermodynamic state, and radiative balance is one of the most challenging problems
We use statistical inversion and machine learning to explore the interaction between changes in the Earths climate system and the radiative fluxes, circulation, and precipitation generated by large scale organized cloud systems.
Prof. Derek Posselt (Atmospheric Oceanic & Space Sciences)
Data+
Inference +
Physical model+
Machine Learning+
Theoretical insight +
Problem-specific thought process
+Computer science
= Useful solution
Data &Comp. Science
High perf. computing
Physics & Modeling