Answers are grounded in this paper and cite their sources.
Machine learning tools are increasingly appearing in chemistry courses, where students use models that rank catalysts, materials, or reaction conditions. Because these models produce numerical predictions, students may assume that any prediction generated by a model is reliable. However, the domain of applicability, or the conditions under which a model can be trusted, is rarely addressed explicitly in instructional settings. This Tutorial presents a structured instructional module that is designed to allow students to examine how model predictions depend on data context and scoring assumptions. The activity uses a fully reproducible synthetic data set representing catalyst screening conditions for hydrogen evolution and oxygen evolution reactions. Students analyze how multiple outcome measures are combined into a single reference score using transparent scaling rules and weights. They then compare model behavior under two scenarios: evaluation within a familiar reaction domain and application of the same model in a different reaction context. A simple domain-of-applicability rule is introduced to illustrate how prediction coverage and prediction correctness may diverge. The materials are organized as a modular activity that can be implemented in two classroom sessions. Precomputed data sets, instructor handouts, and student worksheets allow the tutorial to be used without prior coding experience when the provided outputs are used, although optional scripts are included for instructors who wish to generate the results independently. The module is intended for upper-division undergraduate or beginning graduate courses that introduce machine learning concepts in chemistry.
No takes yet. Share an insight, caveat, or question.
Dalal et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: