NPPS Leadership Team

[Brett Viren <bv AT bnl.gov>]

Hi,

This year there is a Type A LDRD invitation related to AI/ML.  Maybe
you've seen it already but I copy the relevant page from the
announcement's appendix below.

I think PanDA-related development can be made to fit into #3.  If
fruitful we can think about some tie-in with Wire-Cell Toolkit's GPU
service pattern.  I don't have bandwidth to lead any thing but I can
participate, especially if WCT is an ingredient.

I think there may also be potential to propose use of PanDA-based
metaparameter optimization feature but I think it will need invention of
a "payload" job that uses AI/ML to search design "space" for some
optimal detector design.  I don't know how that would look but maybe
someone has clever ideas?

-Brett.


Discovery Science Driven by Human-AI-Facility Integration

BNL, as a Lab at the forefront of experimental facilities design and
Artificial Intelligence (AI) research, has developed the concept of
discovery science driven by the deep integration of humans, AI, and
facilities. The idea is that this integration allows the whole to “think
scientifically,” incorporating scientific goals into the operation of
the facilities and enabling new discoveries. The BNL vision for
human-AI-facility integration is to automate routine processes and to
create layers of nested, intelligent systems that can strategize
together with their users and operators about the best experimental
design, execution, and analysis.

Such a radical shift in facilities design and operation will necessitate
changes to the design of facilities at all levels – accelerators, choice
of technique, detectors, sensors, controls, data acquisition, and
analysis – and an integrated data infrastructure to support this
vision. BNL LDRD investment will therefore focus both on pushing the
boundaries of the state of the art in various domains and on overcoming
operational deployment challenges in highly constrained
environments. Initial research topics are: 1) AI enhanced detectors,
accelerators, and sensors; 2) optimal experimental design and steering;
and 3) migration to operation.

1) AI enhanced Detectors, Accelerators and Sensors

Against the backdrop of exponentially increasing data rates, sustainable
data management will require increasingly powerful, lower-level
processing capabilities to produce the best possible value of
information per available power as a critical figure of merit. These
pipelines will require careful design, management, and optimization, in
terms of placement of compute power, storage, algorithms (including AI),
and network connectivity in accelerators, detectors, controllers,
instruments, and sensors. However, the integration of AI ready advanced
computing capabilities – Neuromorphic, Specialist AI, Graphical
Processing Units (GPU) or Tensor Processing Units (TPU) – is
particularly challenging in harsh experimental conditions as are found
in many experimental facilities.

2) Optimal experimental design and steering

In the context of human-AI-facility integration, a foundation of AI and
Applied Math methods is needed to analyze large, streaming data sets,
integrate information, offer actionable information to users, operators,
and systems to optimize experiments and indeed the choice of experiments
to be performed.  Specific areas of emphasis include continued
development of algorithms for end-to-end automation, uncertainty
propagation, and optimal design under uncertainty for multi-stage,
multi-fidelity, multi- modal workflows, computer simulations (digital
twins) for accelerators, detectors, sensors, controls, experiments,
computational hardware, and data acquisition and analysis, to support
the design and execution of experiments, guide the operation
optimization of facilities, and study their behavior.

3) Migration to Operation

Success in this area will depend on an integrated data infrastructure
fabric that provide seamless and transparent access to AI and HPC
compute, storage, and network capabilities. The user should neither know
nor care where these resources are located and the user experience
should not change if the user is on site or remote. To this end, there
is a critical need for research in: 1) scalable and scientific workflows
with explicit quality of service and performance bounds; and 2) dynamic
end-to-end co- allocation and provisioning of compute, network, and data
resources. POC: Frank Alexander, falexander AT bnl.gov.

Attachment: signature.asc
Description: PGP signature