Logo do repositório

Percorrer por autor "Josefovicz, Matheus"

A mostrar 1 - 1 de 1
  • Gaussian processes in the description of aqueous electrolyte solutions
    Publication . Josefovicz, Matheus; Pinho, Simão; Abranches, João Dinis Oliveira; Watanabe , Erica Roberta Lovo da Rocha
    This work investigates the use of Gaussian Processes (GPs), a Bayesian machine learning method, for predicting osmotic (OC) and activity coefficients (AC) of electrolytes in aqueous solution at 298.15 K. A comprehensive database was constructed from recommended data reported in the literature, containing osmotic and activity coefficients for 1:1, 1:2, and 2:1 electrolytes over a wide range of molalities. The descriptors employed consist exclusively of sigma profiles of cations and anions, generated from COSMO calculations based on Density Functional Theory (DFT), which are combined with electrolyte molality in the first model and with ionic strength in the second model. After a rigorous data curation step that excluded ions affected by the use of standard ionic radii in COSMOtherm, the final dataset comprised 82 electrolytes and 45 distinct ions. The GP models were trained by testing different kernel functions and normalization strategies, with a stratified division of the data into training and validation sets, while the test set was specifically defined to contain representative electrolytes from the dataset. Predictive performance was evaluated using the coefficient of determination (R²) and the Mean Squared Log Error (MSLE). The results demonstrate that GPs can interpolate osmotic and activity coefficients with high accuracy across different classes of electrolytes and molality ranges, achieving R² values close to 1.0 for both the training and validation datasets. As predictive capacity increases, interpolative capacity decreases; however, the model was still able to predict thermodynamic data with good accuracy, achieving R² for test set equal to 0.9013 for OC, and equal to 0.9412 for AC. These findings indicate that the combination of sigma profiles and Gaussian Processes constitutes a promising and computationally efficient approach for predicting thermodynamic properties of aqueous electrolyte solutions, reducing reliance on extensive experimental datasets and offering an interesting alternative to traditional thermodynamic models.
    2026Dissertação de mestrado Acesso restrito Ver mais