Foundations for the future of science discusses using artificial intelligence and machine learning to advance scientific research. Key points discussed include using AI to analyze large datasets, develop scientific models, and automate experimental workflows. The document also outlines several examples of how the Globus data platform is currently enabling AI-powered scientific applications across multiple domains. Overall, the document advocates that embracing "AI for science" has the potential to accelerate scientific discovery by overcoming limitations in human analysis capabilities and computational resources.
5. The COVID’19 data pipeline:
HPC, ML, people developing machine readable datasets for small molecule libraries
CHEMICAL
LIBRARY DATABASE
AND MORE
known
molecules
4B
COMPUTING
RESOURCES
CANONICALIZATION COMPUTE FEATURES DEEP LEARNING
FILTERING
FINGERPRINTING SIMILARITY SEARCH
GENERATE IMAGES CNN FILTERING
Yadu Babuji, Ben Blaiszik, Kyle Chard, Ryan Chard, Ian Foster, Logan Ward, Tom Brettin et al
9. The next frontier? “AI for science”
“Most of the modeling and prediction necessary to
produce the next generation of breakthroughs in
science, energy, medicine, and national security will
come not from applying traditional theory, but from
employing data-driven methods at extreme scale
tightly coupled to experiments and scientific user
facilities.”
— US Department of Energy
FY 2021 Congressional Budget Justification
10. Why am I excited about “AI for science”?
Push
• Step changes in AI/ML
methods, notably deep
neural networks
• Major advances in areas like
machine translation, speech
recognition, image
processing
• New hardware specialized
for deep neural networks
Pull
• Exploding volumes of data due
to new sensors and
instrumentation exceed
human capabilities
• End of Moore’s Law puts hard
problems out of reach
• Growing complexity of science
and engineering problems
slowing rate of discovery
Why are we excited about “AI for science”?
Push
• Step changes in AI/ML
methods, notably deep neural
networks
• Major advances in areas like
machine translation, speech
recognition, image processing
• New hardware specialized for
deep neural networks
Pull
• Exploding volumes of data due
to new sensors and
instrumentation exceed
human capabilities
• End of Moore’s Law puts hard
problems out of reach
• Growing complexity of science
and engineering problems
slowing rate of discovery
11. AI Enabled
Experimental Workflows
(how to make it)
…materials, polymers, organisms…
…self-driving labs, synthesis search…
• data Sets
• literature
• science “news”
• strategy
Cleaned
Updated
Annotated
Aggregated
Interpreted
AI Enabled
Scientific Comprehension
(what it means)
AI-Enabled
Design Workflows
(what to make)
Insight?
AI Science Applications: One per Planet
12. Augmented
Simulations
Design Control
Science and Math
Comprehension
Generative
Models
Inverse
Problems
Multimodal
Learning
Decision-
Making
Materials
Biology
Chemistry
Devices
Batteries
Drugs
Waveforms
Text
Images
Structured
Graphs
Time-
series
Image2Phase
Spectra
2
Structures
Waveform
2
Source
Detector
Simulations
Cosmology
Biodesign
Experiments
Accelerators
Reactors
Mobility
Simulation
Energy
Landscape
Search
Surrogates
Optimize
Mathematics
Physics
Biochemistry
Risk
Assessment
Research
Priorities
The
Next
Problem
AI for Science: AI Building Blocks (examples)
13.
14. Protein
engineering
Liquid-handling
robot
SAXS, SA-
XPCS:
8-ID-I Beamline
Digital twin +
AI components
Robotic
pendant drop
Screen ~108
conditions for
LLPS
Screen ~104 combos for
LLPS (turbidity, confocal
microscope imaging)
Screen ~102 combos at
various temperatures
Selected matrixes
(e.g., salt, pH, PEG)
Stock proteins
(different periods,
repeats)
X-ray
Info transfer and control,
demonstrated
Information transfer and control,
not yet demonstrated
Material transfer, not yet
demonstrated
Change sample Measure sample
HPC simulation
Compute ~105 properties
ALCF
APCF APS
Arvind Ramanathan et al.
Example: Rational design of intrinsically disordered
polypeptides
15. AI for science means rethinking infrastructure
15
Infrastructure for AI-enabled Science
Scientific instruments
Major user facilities
Laboratories
Automated labs
…
Sensors
Environmental
Laboratories
Mobile
…
Simulation codes
Computational results
Function memoization
…
Databases
Reference data
Experimental data
Computed properties
Scientific literature
…
Scientists, engineers
Expert input
Goal setting
…
Industry, academia
New methods
Open source codes
AI accelerators
…
Data
ingest
Inference
HPO
Data
enhancement
Data
QA/QC
Feature
selection
Model
training
UQ
Model
reduction
Active/
reinforcement
learning
Artificial Intelligence
Methods
Data
Models
Accelerators
Compute
Agile
Infrastructure
Surrogates
System Software
Data
mgmt
Operating
system
Portability
Compilers
Runtime
system
Workflow
Automation
Prog.
envs.
Languages
Model
creation
Libraries
Resource
mgmt
Authen/Access
17. Understanding SARS-CoV-2 Protein Structure
17
“These data services have taken the time
to solve a structure from weeks to days
and now to hours”
Darren Sherrell, SBC beamline scientist
APS Sector 19
18. Data Management at Cyro-EM Facilities
18
Case Western Reserve – Cryo-EM Core
Credit: https://case.edu/medicine/research/som-core-
facilities/cryo-electron-microscopy-core
Credit: https://pncc.labworks.org/about-us
Pacific Northwest Cryo-EM Processing Center
(PNNL and Oregon Health Sciences University)
Globus for
– automated data
sync as new data is
collected
– provisioning of data
access for
researchers
– reliable, secure data
access for users
– Monitoring and
management via
console
19. The Bioinformatics Core of the
Lineberger Comprehensive Cancer Center
at the University of North Carolina
Global data distribution at bioinformatics core
– Multiple research projects use Globus for data sharing with external
collaborators
– Support wide variety of projects: different locations, sources, sizes,
cancer types, institution type, storage systems, and identities
20. Digital agriculture – University of Winnipeg
• Increasing crop yields using
machine learning models
• Building training data sets
– 40K images per day, tagged
with metadata
– Move data from diverse
sources to campus storage,
then onto Compute Canada
HPC to run models
• Orchestrate data transfer
using Globus CLI
20
Credit: Dilbarjot and Michael Beck,
Physics and Applied Computer Science , University of Winnipeg
21. Dark Energy Science Collaboration
• Preparation for the arrival
of the Rubin Observatory
• Data Challenge 2:
extreme-scale simulation
of 300 sq degree patch of
the sky over five years
– 5 TB of data
– ~90M core house at ALCF
and NERSC
• Data Portal based on
Globus makes data
accessible to collaborators
21
22. Federated Research Data Repository
• National Research Data
Management platform, where
data can be
– Ingested, curated, and preserved
– Discovered, cited, and shared
• Globus Services
– Authentication
– Transfer to repository service
– Search for metadata catalog for
data discovery (includes metadata
from 70 other repositories)
22
23. Rebuilding A Kidney
GPCR
GUDMAP
Synapse
FaceBase
● DERIVA is an asset management
platform for science used in various
biomedical data repositories
● Globus Auth for authentication with
external identities
● Globus groups for roles (e.g., curator,
viewer, administrator)
● Globus Auth for desktop GUI and CLI
DERIVA
24. 24
Data Provider Models / Functions
API layer
API layer
Data Publishers Model Publishers
Consumers
Science!
Increasing Data Interoperability & Reusability
From foundry import Foundry
f = Foundry()
X,y = f.load(“dataset1”, v=“1.0”)
y_pred = f.run(“model1”, v=“1.0”, X)
f.data.publish(“./”
“dataset1”, v=“1.1”)
f.model.publish(“./”
“model1”, v=“1.1”)
• Models run locally or on distributed endpoints
• Capabilities to pull datasets to desired location
or move compute to desired location
Dataset Function
CH MaD
• Radically reduce the energy barrier to access curated
ML datasets and ML models
• Facilitate reuse, meta-studies, benchmarking, and more
• Long term implications for education
NSF CSSI Started Oct. 2019
(Dane Morgan, Paul Voyles, Michael Ferris, Marcus Schwarting, Ben Blaiszik)
26. Enabled by the Globus data platform
Researcher initiates
transfer request; or
requested automatically
by script, science
gateway
1
Instrument
Compute Facility
Globus transfers files
reliably, securely
2
Globus controls
access to shared
files on existing
storage; no need
to move files to
cloud storage!
4
Researcher
selects files to
share, selects
user or group,
and sets access
permissions
3
Collaborator logs in to
Globus and accesses
shared files; no local
account required;
download via Globus
5
Automating research
workflows and
ensuring those that
need access to the
data have it.
8
Personal Computer
Transfer
Share
• Use a Web browser or
platform services
• Access any storage
• Use an existing identity
Build
The Globus
Command Line
Interface, API sets,
and Python SDK
provide a platform…
6
… for building
science gateways,
portals and
publication services.
7
31. 31
Globus automation services
Managed, secure and reliable task
orchestration
across heterogenous resources,
with declarative language for composition,
and extensible to plugin custom actions,
supporting an event driven execution model,
for automation at scale
32. Create and deploy flows
32
• Define the flow and
deploy to Flows service
• Uses declarative
language (JSON or
YAML)
• Set policy: visibility,
runnable by
Action 1 Action 2 Action 3 Action 4
Action 1
Action 2
Choice
Action 4 Action 5
Action 3
33. Start and manage runs
33
• An instance of Flow
execution
– Provide input parameter
– Check status
– Cancel
• Set policy: monitor,
manager
• Triggers to start flows
34. Build action providers
34
• Action Provider is a
service endpoint
– Run
– Status
– Cancel
– Release
– Resume
• Action Provider Toolkit
action-provider-
tools.readthedocs.io/en/latest
Search
Transfer
Notification
ACLs Identifier
Delete
Ingest
User
Form
Describe Xtract
funcX Web
Form
Custom built
Globus Provided
35. Automation services ecosystem
GET /provider_url/
POST /provider_url/run
GET /provider_url/action_id/status
GET /provider_url/action_id/cancel
GET /provider_url/action_id/status
Create Action
Providers
Define and
deploy flows
{ “StartAt”: ”ToProject”,
”States” : {
”ToProject” : { … },
”SetPermission” : { …},
“ProcessData” : { … } … }}
Run flows
42. Partnership with the community
to develop new connectors
Community Connector Program
43. Easy egress and ingress of data
Data sharing with collaborators
Publish data
44. POSIX Staging Connector
• For POSIX file system that cache from tertiary storage
• Custom plug-in for staging files
• Example:
– IBM Spectrum Scale plugin, Brock Palen at University of
Michigan - github.com/brockpalen/ltfsee-globus
44
46. Globus Groups
• Groups platform in production
• Administrators can add users, in addition to invite
• Membership policies simplified
groups.api.globus.org/redoc
46
47. Transfer and Sharing
• Skip files with not found errors
– List of skipped files once task is completed
• Fail tasks with quota errors
• Scheduled and replicated transfers
– Manage scheduled/repeated transfer and sync tasks
– pypi.org/project/globus-timer-cli
47
53. Globus Connect
• Tools to migrate from v4 to v5
– Migration in phases (Q2 – Q3)
– Goal: not require end user intervention
• IPv6 support
• Connectors
– Azure Blob
– Intel DAOS
53
54. IAM and Data platform
• Support use cases that need higher task throughput
• Enhancements to data permissions management
• Improvements to consent management
• Integration with NIH Researcher Auth Service
• Search service for high assurance tier
• Leverage Search for Globus resources
54
55. Automation platform
• Lower the barrier for adoption
– Web interfaces
– Supporting tools/libraries
– Action Providers for all Globus functionality
• Exemplar flows for common use cases
– Instrument data management
– Data publication
• Supported in high assurance tier
55
56. Clients
• Streamline SDK/CLI to across services
• Web App
– Updated management console
– Accessibility standards
• Enhancements to sample portal
– Open source, for customization and deployment
– Flask, Django
56
59. Requirements for reliable, scalable, remote
computing
Researcher needs to
run a computation on
a remote PC, cloud,
supercomputer
1. Compute
Compute Facility
Collaborator
wants to run their
colleague’s
computation on
another system
closer to their data
3. Share
Instrument
5. Build
Gateway and application
developers want to add remote
computation to their code
2. Specialize
Researcher
needs to move it
to a new system
or architecture to
improve
performance
4. Community
Access
Collaborators
want to share
access to a
single allocation
to run compute
tasks
60. Function as a Service (FaaS)
Developers work in terms of
programming functions
1. Pick a runtime (e.g.,
Python)
2. Register function code
3. Run (and scale)
Low latency, on-demand,
elastic scaling, easy to
deploy and update
60
def compute(input_args):
# do something
return results
61. funcX: managed and federated FaaS
• Cloud-hosted service for managing compute
• Register and share compute endpoints
• Register and share Python functions
• Reliably, scalable, securely execute functions on
remote endpoints
• Integrated with Globus Auth and data ecosystem
61
Try funcx on Binder
https://funcx.org
62. Transform laptops, clusters, clouds into function
serving endpoints
• Python-based agent and pip
installable locally or in Conda
• Elastically provisions resources
from local, cluster, or cloud system
• Manages concurrent execution on
provisioned resources
• Optionally manages execution in
Docker, Singularity, Shifter
containers
• Share endpoints with collaborators
62
$ pip install funcx-endpoint
$ funcx-endpoint configure myep
$ funcx-endpoint start myep
63. Register and share functions
Create funcX client (and authn)
63
def compute(input_args):
# do something
return results
def compute(input_args):
# do something
return results
def compute(input_args):
# do something
return results
Define and register Python function
64. Execute tasks on any accessible endpoint
Select: function ID, endpoint ID, and
input arguments
Retrieve results asynchronously
(funcX stores results in the cloud)
64
F(ep1,1)
F(ep1, 2)
F(ep1, 3)
F(ep1, 4)
F(ep1, 5)
F(ep1, 6)
F(ep2, 7)
66. Canonical research automation flow for instruments
69
Data Capture Data Analysis /
Model in the Loop
Publication
Data Staging
Metadata Extraction
And Data Cataloging
Data Staging
Catalog
Feedback
Data
Generation
Examples
• Serial X-Ray
Crystallography
• X-Ray Photon
Correlated
Spectroscopy
• High energy
diffraction
microscopy
• High throughput
ptychography
• High energy x-ray
diffractions
67. Applying the Globus
platform to science at
the APS
70
Advanced
Photon
Source
Key: funcX agent
Globus Connect
Theta
Bebop
Cluster
Argonne
Leadership
Computing
Facility
Laboratory
Computing
Research Center
Petrel store
APS
Computing
Orthros Cluster
APS DM system
Portal
server
Portal
server
Cooley
Action 1 Action 2 Action 3 Action 4
68. Example: Rapid Training of Deep Neural Networks
using Remote Resources
• DNN at the edge for fast
processing, filtering, QC
• Requires tight coupling
with simulation and
training with real-time data
• Globus Flow:
71
Data Source HPC/DCAI Edge(Host)
Globus,
Automate
Commands
Status
Data Model
User
Request
Status
Commands
Status
C/S
Zhengchun Liu, Jana Thayar, et al.
– Globus to rapidly move data for training
– funcX for simulation and model training
– Globus to move models to the edge
– (Future) funcX for inference at the edge
85. A word from our Platinum Sponsor
Jordan Winkelman, Field Solutions CTO
89
86. Join us in Gather.Town
• Get answers at the Globus Genius Bar
• Visit the Sponsor Showcase
• Joint the scavenger hunt in The Garden
• Play a game
bit.ly/globustown
(passcode: globus)
90