Presentation #541.10 in the session “Computational Augmentation to Surveys and Science Programs”.
We believe that Artificial Intelligence (AI) has potential to be a valuable augmentation to the current processes of the National Academies’ Decadal Survey prioritization of science goals for NASA, NSF, and the Department of Energy. Here we summarize the status of our initial investigation into this potential to identify candidate astrophysics priorities for the 2021+ period, using commercially available AI trained on recommendations in the Astro2010 Decadal Survey. Specifically, we apply an approach based on Latent Dirichlet Allocation and Natural Language Processing to reveal trends in published research that may indicate research priorities. Topic models for ~550,000 digital text abstracts taken from the Astrophysics Data System (ADS) were created for the years 1997 –2010, which we assume was the major body of research underlying the Astro2010 recommended science priorities. We can calculate the growth or decline of 400 identified topic areas over this time period. As a training set, we assess how these topics are related (or not) to science priorities in Astro2010, as well as to the ~300 white papers solicited as part of that Survey. Our assumption is that the growth rate of papers published in a topic area is a good proxy measure for importance of an area of research to the community. There appears to be good, albeit not perfect, correspondence between our machine-learning technique and the recommendations for future high-priority research reported in Astro2010. Differences can also be identified between the apparent recommendations by our AI-based process and those within the solicited white papers. When our training process is complete, our process will be applied to ADS abstracts from 2006 – 2019, which we assume will be the major body of work to underlie national astrophysics science priorities for the period 2021+.