Data Sharing
Data sharing is the practice of making the data behind a publication available in a form that others can find, read and reuse. Funding agencies increasingly require data management plans, but they usually leave the means of sharing to the researcher, and conventional manuscripts convey data mainly as figures and tables that are not machine readable. Data sharing is closely tied to reproducibility and, in the Kitchin group’s practice, to writing papers as executable documents (literate programming).
Data embedded in Supporting Information (2015)
Kitchin describes two complementary ways to share data within the traditional journal workflow (Kitchin 2015). The first attaches data files to the Supporting Information PDF using the standard PDF attachment feature, which any author can do with PDF editors, the LaTeX attachfile package or Word; the group has attached JSON (all DFT geometries and parameters in boes2015a), CSV (curnan2014) and Excel files (hallenbeck2013). The second embeds data as named tables in the org-mode source of the Supporting Information, which is itself attached to the PDF, so the data is documented in place and can be parsed by any scripting language. Very large data sets (about 1.8 GB of DFT calculations for xu2015) are better deposited in a repository that assigns a DOI, such as Zenodo. The paper argues against waiting for standard formats or centralized catalysis databases: authors should share now in fit-for-purpose formats, and domain experts can later curate and convert distributed data. Summary: kitchin2015.
Data sharing beyond atomic positions in Surface Science (2016)
Writing for the Surface Science community, Kitchin argues that the journal’s 2010 policy requiring accessible atomic positions (and the ensuing debate over CIF as a standard) addresses only part of a broader need to share data and methods (Kitchin 2016). Using the JSON file attached to the Supporting Information of a Cu-Pd core-level shift paper (boes2015), he shows that a few lines of Python with the ASE database module can list the stored calculations, write a chosen structure out as a CIF file, and recover the VASP parameters, k-points, cell, positions and total energy needed to rerun it. A change in the ASE database interface broke the originally published code but not the data, which remained readable plain JSON; the paper uses this to argue for plain, documented formats and notes that embedded data in published manuscripts is likely to outlast external repositories, which are still recommended for large data sets (about 1.8 GB on Zenodo for xu2015). Summary: kitchin2016a.
Archiving high-throughput data (2016)
In discussing the outlook for high-throughput methods, Kitchin and Gellman note that both experimental and computational high-throughput work produce datasets too large for standard publications to archive, and point to repositories such as Zenodo (where the group archived a 1.8 GB DFT dataset), Figshare, institutional repositories and the Materials Project (Kitchin 2016). They also argue that the value of such data depends on trustworthy tools for modeling, fitting and visualization, and on training practitioners to use them. Summary: kitchin2016.