QSPR Prediction of Coating Agent Migration Resistance in Nitrocellulose Matrices Based on Bayesian Optimisation and Machine Learning

ABSTRACT

To break the engineering bottleneck of predicting coating‐agent migration resistance in nitrocellulose (NC), this work constructs a small‐sample QSPR model. A dataset of 136 log ₁₀ D records covering 34 guest molecules (plasticizers, energetic additives, and polymer coatings) was built, with 25 topological and physicochemical descriptors extracted from SMILES via RDKit and combined with environmental variables. Under leave‐one‐out cross‐validation (LOO‐CV), raw GPR ( R ² = 0.865, MAE = 0.241) outperformed raw RF ( R ² = 0.813, MAE = 0.460) owing to its smooth‐kernel Bayesian prior, whereas Bayesian optimisation (TPE) barely improved GPR but lifted RF to R ² = 0.897 (MAE = 0.248), whose piecewise structure better captures discontinuous diffusion changes. To assess extrapolation to unseen molecules, grouped Leave‐One‐Molecule‐Out cross‐validation (LOMO‐CV) was implemented: RF + BO attains R ² = 0.397 versus 0.107 for GPR + BO, revealing that interpolation metrics substantially overestimate extrapolation capability—a caveat we report explicitly. Against traditional baselines, RF + BO matches per‐molecule Arrhenius fits in interpolation (0.897 vs. 0.893) and outperforms the Fedors group‐contribution method by ∼30‐fold in cross‐molecule prediction (0.397 vs. 0.013). Feature importance analysis highlights hydrogen‐bond donors (23.9%), molecular weight (13.8%) and molecular surface area (11.1%) as dominant factors, verifying interfacial hydrogen bonding and steric hindrance as key anti‐migration mechanisms. This workflow delivers a unified QSPR framework for screening low‐migration coating agents.