Predicting Critical Parameters of Hydrocarbon Species using Bidirectional Encoder Representations from Transformers

The development of high-performance combustors operating at supercritical chamber pressures necessitates high-fidelity simulations that accurately account for real-fluid effects. However, the reliability of real-fluid equations of state (EOS), such as the Peng-Robinson model, is frequently compromised by the scarcity of experimental critical property data for short-lived intermediate species. This study presents a novel machine learning framework for predicting critical temperatures ($T_c$) and critical pressures ($P_c$) by fine-tuning a pre-trained Bidirectional Encoder Representations from Transformers (BERT) model. Utilizing Simplified Molecular Input Line Entry System (SMILES) strings for molecular representation, the model exploits attention mechanisms to learn complex structure-property relationships from a dataset of hydrocarbon species. A sensitivity analysis conducted on a counter-flow diffusion flame reveals that thermodynamic transport properties (specifically viscosity and density) are highly sensitive to perturbations in critical parameters, underscoring the necessity for high-precision estimation methods. The proposed SMILES-BERT architecture demonstrates robust performance, achieving low prediction errors on unseen test data and exhibiting strong generalization capabilities without overfitting. This approach offers a scalable, data-driven solution for generating the critical parameters required to enhance the accuracy of computational fluid dynamics (CFD) simulations in transcritical regimes.