Concrete research audit finds full data for 90 of 419 studies
A peer-reviewed audit of machine-learning research on concrete found complete datasets for 90 of 419 papers and source code for 20. Researchers checked access to the files needed for computational reproduction, separating original-data citations from the actual training tables. They assessed sharing practices without rerunning the models or checking the correctness of their results.
Science··Morning
Full datasets were available for 21.5 percent of papers
A peer-reviewed audit by researchers Jeonghyun Kim and Diana Bajare found that most sampled papers predicting concrete properties with machine learning did not make their computational assets fully accessible. Complete datasets could be obtained for 90 of the 419 studies, while actual source code was accessible for 20. The investigation examined the availability of research files, without executing the models or judging their numerical accuracy.[1]
The dataset categories included 211 papers offering only a citation to the original data source and another 118 sharing no data. These represented 50.4 percent and 28.2 percent of the sample. Full dataset availability stood at 21.5 percent, compared with 4.8 percent for source code. Of the remaining code categories, 380 papers shared none, 17 provided pseudocode and two had nonworking links.[1]
Original-data citations differed from training tables
The audit distinguished a citation to measurements collected elsewhere from access to the assembled, preprocessed table actually used to train a particular model. Combining source datasets can leave researchers with a new training table. The classification therefore tracked access to that working dataset separately from whether an article named an earlier data source.[1]
The sample came from a June 13 search of the Science Citation Index Expanded within Web of Science, a research-publication index. A 2025 publication-year filter initially returned 838 articles written in English. The retained studies used tabulated concrete-mixture variables and excluded other tasks, including image-based crack detection. The original search membership was preserved, so 38 entries subsequently dated 2026 remained among the 419 papers.[1]
A second assessor checked 84 classifications
A second assessor independently classified 84 randomly selected papers. Agreement reached 96.4 percent for datasets, 97.6 percent for code and 98.8 percent for graphical interfaces. Among 85 studies that developed an interface, 52 restricted access. The researchers did not require an interface for computational reproduction; they examined it as a separate research asset.[1]
The authors recommend archiving versioned datasets, code, dependency information and persistent links together. Their audit covered one index and one research field; it did not evaluate file reuse, licensing or complete compliance with data-management principles. Missing access was not classified as evidence of misconduct. The statistical interval for full-data availability was 17.8 percent to 25.7 percent at 95 percent confidence. The corresponding code-access interval ran from 3.1 percent to 7.3 percent.[1]
Related columns
For more information on this topic, you can read the related columns.