Automated High-Throughput Experimentation (HTE) workflows are increasingly used in catalysis to generate large and reliable databases of Quantitative Structure-Properties Relations (QSPR). Data-driven approaches integrating HTE and Artificial Intelligence (AI) tools such as Machine Learning (ML) and Deep Learning (DL), can be exploited to rapidly and thoroughly navigate complex variable hyperspaces and build models predicting catalyst performance. In a recent publication we highlighted the utilization of a custom-made HTE/AI workflow for the preparation, screening, and “black-box” QSPR modeling of a large library of “High-Yield” Ziegler–Natta (HY-ZN) catalyst formulations, with the ultimate goal of identifying Internal Donors (ID) specifically for tunable applications. In the present paper, we illustrate how a smaller but more homogeneous ID subset containing diesters only can be utilized for “clear-box” QSPR modeling also aiming at increased mechanistic insights. The study led to unconventional conclusions that challenge some long-standing hypotheses about the role of surface modification by electron donors in HY-ZN catalysis. In particular, evidence was achieved that the ID leaves a permanent footprint in the catalyst, which durably affects catalyst performance even in case the ID is reactive with the AlEt3 activator and is extensively removed from the solid phase during polymerization.