Evidence dataset for Machine Learning for Reactive Transport Modelling: A Critical Review of Evidence for Physical and Chemical Consistency

This dataset accompanies the manuscript "Machine Learning for Reactive Transport Modelling: A Critical Review of Evidence for Physical and Chemical Consistency" by Ramin Mahyapour and Mahmood Sadat-Noori. It contains the curated publication register, operational indicator definitions, source-based reporting decisions, quantitative extractions, source locators and available selection records supporting the review. The assessed set contains 32 publications, including 28 chemically eligible publications used for reporting frequencies. The workbook contains 320 study-indicator decisions, 24 worksheets, 958 formulas and 18 native tables. Version 3.0 retains all codes and publication membership from version 2.0. Table 3 now reports Y / non-NA, with non-NA defined as Y + NR + U, in place of Y / 28. The new Table S8 separates SHIFT test types and OBS comparison types, with formula-linked study assignments and source details. Five non-exclusive SHIFT groups contain 3 within-family or sampled-design contrasts, 2 parameter-range extrapolations, 3 temporal extrapolations, 2 spatial-discretization or cross-geometry transfers, and 2 boundary-condition or laboratory-core transfers. These 12 assignments cover 10 publications. Five descriptive OBS groups contain 2 held-out laboratory-core studies, 1 randomly held-out measurement-point study, 1 laboratory comparison of a simulation-trained surrogate, 1 fitting-informed comparison, and 1 combined withheld-variable and process-model evaluation. The two laboratory-core publications share a dataset. The README explains identifiers, codes, calculations, changes and reuse. Quantitative results for all 28 eligible publications, primary-source locations, modelling roles, uncertainty targets and reporting frequencies by modelling role remain available in the worksheets introduced in version 2.0. The data support a publication-level assessment of reporting within a curated evidence set. A qualifying report is distinct from test success or independent replication. Source-inspection records retain their documented scope; independent duplicate coding and model re-execution were not performed. The deposit contains extracted results and documentation, not new experimental or simulation outputs.