Abstract
This study evaluates machine learning (ML) methodologies in the pursuit of advancing thermodynamic flash calculations that are vital for carbon dioxide storage applications and for the oil and gas industry in general. We generated a dataset for training machine learning algorithms using a traditional physics-based model. This developed hybrid model incorporated into the machine learning model underlying physical constraints. While preliminary results from training and numerical matching were promising, the hybrid model's real-world application revealed non-trivial shortcomings. Specifically, mismatch in the multiphase region was observed during compositional space testing. Such subtle but significant flaws in machine learning methods have profound implications for the accurate physics of carbon storage projects. This article, therefore, presents advantages and disadvantages of employing ML for thermodynamic calculations, emphasizing the intricate balance between computational efficiency and representative physics.