Machine Learning Solubility Estimation via Molecular Descriptors

The patent applies preliminary action by pre-training a machine learning model using experimental solubility data from multiple solutes and solvents before actual solubility estimation is needed. The model is trained in advance on a dataset containing chemical structures and corresponding solubility parameters, so that when new solubility predictions are required, the system can quickly process new queries without conducting new experiments. This pre-computation approach resolves the contradiction by preparing the predictive capability beforehand, eliminating the need for time-consuming repeated experiments while maintaining accurate solubility data through the trained model.

The patent applies copying by using a machine learning model to replicate the relationship between chemical structures and solubility parameters that was originally established through experimental data. Instead of physically repeating experiments to obtain solubility data for new solute-solvent combinations, the system creates a computational copy of the experimental knowledge embedded in the trained model. This computational copy can rapidly predict solubility for any new combination by processing chemical structure inputs, thereby avoiding the need for actual repeated experiments while maintaining measurement precision through the model's learned patterns.