Use this URL to cite or link to this record in EThOS:
Title: Application of machine learning methods for design of crystallisation processes
Author: Gurung, Rajesh
ISNI:       0000 0004 8509 5888
Awarding Body: University of Strathclyde
Current Institution: University of Strathclyde
Date of Award: 2020
Availability of Full Text:
Access from EThOS:
Access from Institution:
There is great potential for the implementation of machine learning to aid in pharmaceutical process development. Machine learning (ML) algorithms can be applied to increase the speed with which high-value drug products are developed for the market while reducing the utilisation of material, minimising wastage and assuring the desired quality attributes are achieved. This thesis illustrates the application of ML techniques in aspects of crystallisation process design assessing the ability to predict crystallisation outcomes, including crystal habit and non-aqueous solubility of pharmaceutical drugs in a diverse range of solvents. High throughput screening for analysing crystallisation outcomes and crystal habit of paracetamol in a diverse range of solvents was developed using Technobis Crystalline. Out of 94 solvents, paracetamol was observed to crystallise in 44 solvents, remain in solution at set conditions in 11 solvents, never solubilise in 36 solvents and show signs of degradation in 3 solvents. Based on these experimental data, a ML classification model was constructed for predicting the crystallisation outcomes and crystal habit of paracetamol with ~77.78 % prediction accuracy. Analysis of the ML model revealed that the physicochemical descriptors and predictive capabilities were more directed towards defining solubility of paracetamol rather than its nucleation behaviour. A rapid and efficient solvent selection tool based on relative solubility was developed using ML algorithms. The tool was not only successful in aid of rational selection of solvent but also reduce the number of screening experiments in the laboratory and thus limit material cost and usage. The regression and classification models built to predict non-aqueous solubility on 247 drug and drug-like molecules in seven commonly used solvents demonstrated that the molecular descriptors calculated using MOE were better at predicting solubility compared to structural fingerprint descriptors. Furthermore, both the regression and classification models successful predicted solubility of drugs in alcohols compared to other organic solvents.
Supervisor: Johnston, Blair ; Florence, Alastair Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral