Giri Krishnan created AIRAVATA-3965:
---------------------------------------
Summary: Facilitating computational experiment generation in
AIRAVATA
Key: AIRAVATA-3965
URL: https://issues.apache.org/jira/browse/AIRAVATA-3965
Project: Airavata
Issue Type: New Feature
Reporter: Giri Krishnan
Computational sciences involve extensive experimentation which often involves
searching over space of parameters, variables, functions and workflows.
Individual researchers and groups often perform a large number of such searches
to identify critical functional forms and workflows for any particular study.
The goal of this work is to provide a tool that facilitates this search
process. This will enable visualization, identifying or learning templates and
generate potential experiments based on past experiments using LLM and
neurosymbolic methods.
This task requires the following specific goals for this work :
# Provide visualization of past computational experiments: Tracking various
computational experiments with various variations is often a challenging
problem for individuals and groups of researchers. Often various adhoc
approaches (directories, git etc) are used to track these changes, but often it
is very difficult to provide an entire overview of past experiments. The goal
of this work is to develop a visualization approach that allows to examine all
the past experiments. This will require dimensionality reduction on the
embeddings from LLMs which have been tested on its code generation abilities
(eg. codellama, Llama 4 Maverick) for generating visualizations. Further
comparison in the performance with standard code cloning and similarity
measures will be required.
# Identify template based on past experiment database: It is common for
several computational experiments to share a common structure, in such cases
identifying the 'template' allows for identifying common approaches in past
experiments and to generate new ones. This work will need software engineer
approach and AI based approaches to identify such templates. The templates will
also be integrated with the visualization (in addition to embedding based
visualization) allowing for examining the collections of experiments that
belong to each template.
# Generate new suggested experiments using templates and visualization guided
search. Generation of new experiments is a key component of computational
science work. To facilate this process, will require a visual interactive way
to generate experiments based within the regions of previous experiments and
also in the space where it was not previously explored. In addition, this will
require generation of new experiments based on templates that were identified
from the previous step. Template based generation could also provide a
verifiable way to generate experiments.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
