Predicting Heat of Combustion from SMILES with Machine Learning

Accurately predicting the heat of combustion of chemical compounds is essential for applications in energy research, material science, and chemical engineering. Traditionally, bomb calorimetry is used to determine this value; however, this method is time-consuming, costly, and poses safety risks. In this study the use of machine learning is explored for heat of combustion prediction of molecules directly from their structural representations. A dataset was compiled from multiple sources, including experimental data from the National Institute of Standards and Technology (NIST) and PubChem, with preprocessing to standardize sample format and remove samples with non-standard data. A neural network was developed and trained using ChemBERTa-generated<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">10</sup> embeddings of molecular SMILES representations. Initial results using a limited dataset produced an R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> value of 0.522 and a mean squared error (MSE) of 9,970,441, suggesting the need for additional data. A synthetic dataset was created by duplicating the existing set. Using this dataset the model’s performance significantly improved, achieving an 80% reduction in MSE and an R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> of 0.896. These results demonstrate the feasibility of using machine learning to approximate combustion energy, reducing reliance on experimental methods. Future work will focus on expanding the dataset, refining hyperparameters, and improving model scalability with GPU acceleration to enhance predictive accuracy further.