Collaborative research: CDS & E: using transients and variable STARS to spur interpretability of foundation models in the rubin lsst ERA

The National Science Foundation Directorate for Mathematical and Physical Sciences awarded $577,695 to the Research Foundation of the City University of New York (doing business as RFCUNY - Hunter College) on September 1, 2026, under the Mathematical and Physical Sciences program (CFDA 47.049) for collaborative research on foundation model interpretability using transient and variable stars in preparation for the Vera C. Rubin Observatory's Legacy Survey of Space and Time.

The research team at CUNY Hunter College and Columbia University will conduct formal generalizability and mechanistic interpretability analyses of astronomy foundation models in collaboration with the FM teams at the NSF-Simons AI Institute SKAI. The work tests whether cutting-edge astronomy foundation models outperform narrow, task-specific machine-learning models when processing large-scale data from LSST, and examines how these models learn to generalize. The team will participate in PROJECT BUILD, which provides tiered mentoring to support undergraduate researchers. This award is partially funded by a gift from Charles Simonyi to the NSF Directorate for Mathematical and Physical Sciences Section of Astronomical Sciences.

Performance occurs at the recipient's location in New York, NY 10065. The period of performance runs from September 1, 2026 through August 31, 2029. The assistance type is a Project Grant.

Abstract

In science, the deeper a physical truth, the more widely IT applies. A similar principle is thought to hold for artificial intelligence (AI) and machine learning (ML)?the more deeply a model has ?learned?, the more generalizable IT is. Generalizability therefore sits at the heart of AI/ML. To test and improve astronomy foundation models (FMS), a research team at cuny hunter college and columbia university will collaborate with the FM teams at the NSF-simons AI institute skai to conduct formal generalizability and mechanistic interpretability analyses. The plan is to assess the generalizability of some of the most cutting-edge types of models being adopted by the astrophysics community and to study how they learn to generalize, thus paving the way for their effective, reliable, and ethical use. The team will also join project build, which is modeling a tiered mentoring system to provide a supportive research environment for participating undergraduates. This research award is partially funded by a generous gift from charles simonyi to the U.S. NSF directorate for mathematical and physical sciences section of astronomical sciences. The project includes significant contributions to vera C. Rubin observatory?S legacy survey of space and time. The research goal of this project is two-fold. Using quantitative measures of performance and new interpretability methods, the research team will test the contention that astronomy foundation models out-perform?or have the potential to out-perform?narrow, task-specific machine-learning models in the ERA of large-scale data, such as that produced by the vera C. Rubin observatory legacy survey of space and time (LSST). The project will also advance knowledge of how these models actually learn astronomy. The team has designed its research plan so that in the process of accomplishing the first aim, they will either uncover or disprove the existence of a purportedly hidden large population of wide interacting binary STARS of importance for binary stellar evolution, accretion physics, and cosmology. The astronomy science case of searching for wide interacting binary STARS?symbiotic binaries?is extremely well suited for driving the ML[/FM] generalizability investigation because: 1) the phenomenology and underlying physics involved overlaps with and goes beyond the current focus of skai FMS on transients and variable STARS; and 2) characterizing the population of wide interacting binaries is important in its own right. The investigators will use ML analysis with both task-specific and foundation models of skymapper data-release 4, then light curves from LSST alert brokers, and then LSST data-release 1, to reduce uncertainty on the size of the population of symbiotic binaries from several orders of magnitude to less than one order of magnitude. They will also determine whether symbiotic binaries without shell burning on the accreting white dwarf, which currently constitute a small fraction of known symbiotics, actually dominate that class of objects. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the foundation's intellectual merit and broader impacts review criteria.- subawards are not planned for this award.