Modeling Solvatochromic Effect by a Two-Step Machine Learning Approach for Organic Chromophores

Solvatochromism, also known as the environment-induced modulation of photophysical properties, plays a critical role in the design of fluorescent probes, sensing systems, and optical materials. Traditional machine learning models often struggle to disentangle solute-intrinsic effects from solvent-induced shifts, limiting both interpretability and predictive robustness. To address this challenge, we propose a two-step regression framework that models fluorescence-related properties by first estimating a solute-only baseline, followed by a correction step capturing solvent contributions. Utilizing Morgan fingerprints for molecular representation and random forest regression for learning, the model was benchmarked on five comprehensive data sets encompassing wavelength of maximum adsorption and wavelength of maximum emission. Across all data sets, the two-step approach consistently outperforms conventional one-pot models in both accuracy and mean absolute error. For instance, the Q² and MAE of the testing set for predicting λ_emi of the Deep4Chem data set are 0.844 and 27.8 nm, respectively, using a two-step approach, outperforming the corresponding results at 0.767 and 33.1 nm, respectively, using a one-pot model. Permutation importance analysis reveals a chemically intuitive rebalancing of feature influence: solute descriptors dominate prediction, while solvent effects are modeled as secondary perturbations. This modular design not only enhances predictive power but also aligns more closely with physical insights. The framework offers an alternative direction for chemically interpretable, data-driven modeling of the solvatochromic behavior in molecular systems.