Dissertation Defense: Eric Melin

Title: WHEN SCIENCE ADMITS ITS SHORTCUTS: DETECTING AND MANAGING SELF-ADMITTED TECHNICAL DEBT IN SCIENTIFIC SOFTWARE

Program: Computing Ph.D.

Committee Chair: Francesca Spezzano

Committee: Francesca Spezzano, Nasir Eisty, Bogdan Dit

Abstract: Scientific software underpins modern discovery, from climate modeling to computational physics, yet its development often involves acknowledged shortcuts, unresolved assumptions, and temporary compromises. These admissions, known as Self-Admitted Technical Debt (SATD), pose unique risks where software correctness directly affects research validity and reproducibility. While SATD has been studied extensively in general-purpose open-source software (OSS), existing taxonomies, classifiers, and prioritization approaches were not designed for scientific contexts, leaving it unclear how SATD in scientific software should be identified, which debt
receives attention, and how much of it persists. This dissertation investigates SATD in scientific software through three studies spanning identification, prioritization, and resolution. First, scientific debt was formally operationalized and validated with
domain practitioners as a domain-specific form of SATD, and a multi-artifact dataset of 115,524 labeled texts was constructed. A multi-task transformer classifier applied to 900,358 artifacts from 23 scientific software projects showed that SATD is most
prevalent in pull requests and issues, and that a classifier limited to traditional SATD categories misses 74% of scientific debt instances. Because identification alone does not indicate which debt matters, the second study developed an embedding-based prioritization heuristic and traced SATD across linked artifacts in nine projects. Debt in code-proximate artifacts and debt expressed with negative sentiment received
higher priority, and cross-artifact propagation was rare but marked higher priority debt. Only 37.8% of SATD comments were ever removed, and those that were
persisted for a mean of roughly 690 days, both worse than OSS benchmarks. The third study examined what drives this persistence by comparing nine scientific and
nine matched OSS projects using regression, survival models, and static analysis. Resolution was determined primarily by structural factors at introduction, such as
file path type and maturity, rather than by author or sentiment. Documentation and scientific debt showed a compounding resolution disadvantage with age that was not observed in OSS, and SATD in scientific software was introduced into files with 2.25
times the SQALE debt ratio and 2.7 times the code smells of OSS files. Together, these findings demonstrate that SATD in scientific software differs qualitatively from that in general-purpose OSS: it is under-detected by conventional approaches, persistent, and prone to entrenchment. The datasets, models, and empirical evidence presented
here give practitioners and researchers a foundation for domain-aware tracking, monitoring, and time-bounded escalation of SATD, improving the maintainability and reproducibility of scientific software.