Conformer-Resolved Optical Rotation Prediction From Three-Dimensional Quantum and Semiempirical Molecular Fields Using 3D Convolutional Neural Networks.

Optical rotation is highly sensitive to molecular conformation, making prediction expensive when conformational ensembles and electronic-response calculations are required. We evaluate three-dimensional convolutional neural networks trained on volumetric molecular descriptors for conformer-level optical rotation. A chemically structured set of substituted phenylethanol and phenylethylamine derivatives yielded 82,790 valid conformer-level sTD-DFT reference records. Three descriptors were compared on a common grid: HF/STO-3G electron density, B3LYP/cc-pVDZ electron density, and GFN2-xTB charge fields, each combined with atom-type channels. Performance was assessed using random conformer splits, structured chemical holdouts, and an independently generated CREST/GFN2-xTB conformer-ensemble benchmark for the same 65 compounds (705 conformers). Conformer-level predictions were Boltzmann-weighted and compared with sTD-DFT molecular rotations. The GFN2-xTB model gave the best deployment performance (MAE 15.3 deg dm-1 (g mL-1)-1, RMSE 22.2, R2 = 0.768, Pearson r = 0.901, 93.8% sign accuracy). Compound-specific GFN2-xTB models did not outperform the full-dataset model, indicating that shared chemical and conformational information improves transfer. These results show that inexpensive semiempirical charge fields can retain sufficient conformer-resolved chiroptical information for rapid ensemble optical rotation prediction within this chemical series.