Mapping the Flammability Space of Sustainable Refrigerant Mixtures through an Artificial Neural Network Based on Molecular Descriptors

Identifier:  imarina:9378598

Handle: https://hdl.handle.net/20.500.11797/imarina9378598

Authors:  Albà, CG; Alkhatib, III; Vega, LF; Llovell, F

Abstract:
As the EU's mandates to phase out high-GWP refrigerants come into effect, the refrigeration industry is facing a new, unexpected reality: the introduction of more flammable yet environmentally compliant alternatives. This paradigm shift amplifies the need for a rapid, reliable screening methodology to assess the propensity for flammability of emerging fourth generation blends, offering a pragmatic alternative to laborious and time-intensive traditional experimental assessments. In this study, an artificial neural network (ANN) is meticulously constructed, evaluated, and validated to address this emerging challenge by predicting the normalized flammability index (NFI) for an extensive array of pure, binary, and ternary mixtures, reflecting a substantial diversity of compounds like CO2, hydrofluorocarbons (HFCs), hydrofluoroolefins (HFOs), six saturated hydrocarbons (sHCs), hydroolefins (HOs), and others. The optimal configuration ([61 (I) x 14 (HL1) x 24 (HL2) x 1 (O)]) demonstrated a profound fit to the data, with metrics like R-2 of 0.999, root-mean-square error (RMSE) of 0.1735, average absolute relative deviation (AARD)% of 0.8091, and SDav of +/- 0.0434. Exhaustive assessments were conducted to ensure the most efficient architecture without compromising the accuracy. Additionally, the analysis of the standardized residuals (SDR) and applicability domain (AD) exhibited fine control and consistency over the data points. External validation using quaternary mixtures further attested to the model's adaptability and predictive capability. The exploration into the relative contribution of descriptors led to the identification of 23 significant sigma descriptors derived from conductor-like screening model (COSMO), responsible for 90.98% of the total contribution, revealing potential avenues for model simplification without a substantial loss in predictive power. Moreover, the model successfully predicted the behavior of prospective industry-relevant mixtures, reinforcing its reliability and opening the door to experimentation with untested blends. The results collectively manifest the developed ANN's efficiency, robustness, and adaptability in modeling flammability, catering to the demands of industry standards, environmental concerns, and safety requirements.