Interpretable ML-DFT Framework for Performance Prediction and Structure–Activity Relationship Analysis of Acidic Copper Plating Levelers

High-performance levelers are crucial for microelectronic
copper
electroplating; yet, their development is hindered by costly experimental
screening and unclear structure–activity relationships (SAR).
This study proposes an integrated machine learning (ML) and density
functional theory (DFT) framework for the rapid discovery of novel
levelers. First, a data set was established using experimental Dissolution
Peak Decrease Amount (DPDA) values from the literature and DFT-calculated
adsorption energies (Eads) as target values,
with 24 theoretical calculation molecular properties serving as features.
Predictive models, including XGBoost Regression (XGBR) and the Classification
and Regression Tree (CART), were constructed and trained. Using these
models, the DPDA and Eads of 521 effective
moleculesprescreened from a database of 29,785,186 compounds
based on structural criteriawere predicted, leading to the
identification of five superior novel levelers (Res-1 to Res-5). Pearson
correlation and SHAP analyses identified the key descriptors governing
leveling efficacy. Finally, multiangle theoretical evaluations confirmed
that Res-5 significantly outperforms the industrial standard Janus
Green B (JGB). Its exceptional performance originates from the electrophilic
molecular backbone and stable weak interactions between N–N
functional groups and the copper surface. This research not only provides
a high-quality screening tool for high-performance levelers but also
offers new insights into molecular SAR at complex electrochemical
interfaces.