Large language models on tabular data based molecular property prediction (MPP) of RP-3 kerosene

Accurate modeling of multicomponent fuel properties is important for surrogate fuel design, where complex composition-property relationships motivate the development of advanced data-driven prediction strategies. In this study, we evaluate pretrained molecular and general-purpose language model representations for predicting physiochemically distinct properties of RP-3 kerosene surrogates using SMILES-derived embeddings and composition-based information. Four supervised machine learning (ML) models (SVM, GPR, RF and k-NN) were compared with three transformer-based representations (ChemBERTa, MolFormer and TinyLlama (1.1B)) across different dataset configurations. The results demonstrate that the predictive performance depends on both dataset size and feature representation strategy. For the 1000-sample dataset, TinyLlama achieved stable predictive performance (average R 2 = 0.9962), whereas the expanded 6000-sample dataset enabled ChemBERTa to achieve high prediction accuracy, explaining 99% of the property variance (RMSE = 0.126, MRE = 0.18%) outperforming both MolFormer and TinyLlama. Notably, TinyLlama lacking any chemistry specific pretraining, still achieves R 2 = 0.995, purely from serialized composition and structured prompts, demonstrating that contextual language embeddings can encode chemically meaningful information within the investigated RP-3 surrogate fuel system. Cosine similarity and residual analyses confirm that the pretrained representations capture meaningful differences between aromatic and saturated hydrocarbon structures while maintaining strong predictive stability. The controlled ablation study further demonstrates that composition information alone provides highly accurate prediction within the investigated surrogate system, while molecular embeddings provide additional chemically informed representation capability. These findings establish a framework for evaluating pretrained molecular and language-model representations for RP-3 surrogate fuel property prediction, with potential extension toward broader hydrocarbon fuel systems after validation using more diverse molecular libraries and experimental datasets.