Speaker
Description
The Evaluated Nuclear Structure Data File (ENSDF) database is the comprehensive
nuclear structure and decay database, containing all published experimental
nuclear structure and decay data for the 3300 observed nuclides, along with
carefully evaluated recommended values and uncertainties. ENSDF is the leading
source of nuclear structure and decay data worldwide for both basic and applied
sciences, exceeding 3.2M lookups per year via the NNDC website.
The complete modernization of the ENSDF database is underway--an ambitious
project that is transforming the restrictive 80-column ASCII format in use
since its inception into an expansible Object-Oriented Database and API. In
addition to the existing ENSDF data, this upgrade will allow for the addition
of a wide variety of content (such as continuous spectra), efficiently deliver
complex search results, and provide straightforward data access to modern
computational tools. The modernized database will be amenable to the addition
of arbitrary types of open data, as well as supplemental information from
evaluators to speed subsequent evaluations.
One of the bottlenecks in curating nuclear data from the scientific literature
is extracting table contents from PDF (Portable Document Format) documents for
evaluation, particularly because standard PDF table extraction tools produce
many artifacts which are laborious to correct. New machine learning (ML) tools
and techniques are being developed in parallel to automate the insertion of
newly-published data into the companion database for pre-evaluation published
results (XUNDL), which will significantly reduce the labor required of the
compiler. Our approach is to develop machine learning models to 1) isolate
tables in documents, 2) visually detect individual cells, 3) predict table
structure, and 4) recognize the text contents. These machine learning models
are developed using a combination of convolutional and recurrent neural
networks.
The ENSDF data format and storage are being migrated to JSON and the CouchDB
database. This has a number of benefits. Using a standard format and database
management system will greatly reduce the time required to understand the data
structure and encourage community contributions to API development. The new
format will allow ENSDF to exploit existing tools in modern databases and
support rigorous accuracy testing. The data structure will be easily
comprehensible and human-readable for current and future developers, as well as
machine-readable for AI/ML and other applications. The hierarchical key-value
structure of JSON will enable unambiguous direct references to any data in the
upgraded ENSDF, with an intuitive syntax.
Efficient search methods for physical properties such as level half-lives,
$\gamma$-decay energies, etc. are being developed, in parallel with a
Python API. The API will include high-level functions and tools to utilize
the search capabilities. Users will be able to write their own Python scripts
and applications, using the API to access data directly from the new ENSDF
server. The database server will be able to export the majority of the data
currently available in ENSDF to the original 80-column format for legacy codes.
ENSDF is intended to support the research community. As such, user input is
being solicited, and a number of outreach workshops will be held throughout the
duration of the project. New interfaces will be designed for the evaluator and
the user community, and a visualization tool to clarify the data structure is
under development.
This Research is sponsored by Office of Nuclear Physics, Office of Science, US
Department of Energy, under contracts DE-AC02-98CH10946 (BNL) and
DE-AC-06CH11357 (ANL).