Self-solvation energies: Extended open database and GNN-based prediction

Abstract

Solvation energies play a crucial role in various chemical processes, ranging from chemical synthesis to separation techniques. To optimize these processes, it is essential to accurately predict solvation energies across different temperatures and solvents. However, most existing studies primarily focus on the standard temperature of 298.15 K. In this work, we address this limitation by creating an extensive database, which combines the DIPPR and Yaws databases. Our comprehensive dataset includes 5420 pure compounds, resulting in 71,656 data points spanning a wide range of temperatures. Additionally, besides the development of this novel database, another key contribution of this work is the coupling of the well-known Graph Convolutional Neural Network Chemprop, with our database with the aim of predicting self-solvation energies across diverse temperatures for the first time. The results presented here demonstrate the overall effectiveness of the model, evidenced by a low Mean Absolute Error (MAE) of 0.09 kcal mol⁻¹ and a high Determination Coefficient (R²) of 0.992. Additionally, the Average Relative Deviation (ARD) of the data is 2.2 %, further confirming the accuracy of the model. In fact, the model demonstrates robust predictive performance across data of varying quality, including a significant fraction of pseudo-experimental values derived from predictive schemes. However, it is worth noting that some groups of compounds, such as small sized compounds and low-numbered ring structures, exhibited somewhat larger deviations than expected. This suggests areas for further refinement and indicates that while the model is robust, there is still room for improvement in specific cases. This approach represents an overall improvement over previous models and offers enhanced versatility for practical applications in chemical synthesis and separation processes.