Learning to hear: Predicting speed of sound in deep eutectic solvents via molecular signatures

Abstract

Deep eutectic solvents (DESs) are gaining attention as sustainable alternatives to conventional solvents due to their tunable properties and environmental advantages. Yet, predicting their physical behavior, particularly acoustic properties like the speed of sound, remains limited, especially in approaches that consider chemical structure. This study introduces a hybrid machine learning framework that incorporates molecular-level information to predict the speed of sound in DES systems with high accuracy and interpretability. A dataset of 1001 experimental data points, collected from 92 distinct DESs under atmospheric pressure and varying temperatures, was used to train and evaluate the models. Instead of relying solely on bulk experimental parameters, the models integrate structural descriptors based on group and atomic contributions, capturing detailed chemical features of both hydrogen bond donors and acceptors. These structure-based inputs were embedded into machine learning algorithms to develop robust and generalizable predictive tools. The resulting models demonstrated excellent performance, achieving average absolute relative deviations below 1 %. In addition to strong predictive power, the framework allows for interpretation of how specific molecular characteristics influence acoustic behavior, offering a transparent view into the structure–property relationship. This work advances the application of machine learning in solvent science by uniting chemical intuition with data-driven modeling. The hybrid approach not only enhances predictive capability but also provides mechanistic insight, making it a valuable tool for the rational design and selection of DESs in green chemistry, materials processing, and related fields. © 2025 Elsevier B.V.