Automated lab experiments with closed loop optimization via Bayesian optimization (Gaussian process regression) or other ML/AI methods have seen more research traction during the last years for different application domains (e.g. formulation development, drug discovery, solid materials screening) [1,2,3]. Important steps in this closed-loop optimization scheme for the design and execution of experiments are: 1. In-silico feature set generation of solutes and solvents with the COSMO solvation model [4] and other molecule descriptor tools [5]. 2. 2. GPR model settings, initialization and training with in-silico data to generate first design of experiments. 3. Automated assay/formulations generation with laboratory equipment. 4. Automated measurements of solubilities with the generated assay/formulations. 5. Hyperparameters tuning of Gaussian process with measurement data. 6. Evaluation of Gaussian process confidence and end of experiments loop or updated/new design of experiments and execution of the next batch of experiments in the laboratory. We showcase in this work the evaluation results of step 1 and 2 with a thorough analysis how a Gaussian process should be implemented to achieve an efficient high-througput screening study with an automated laboratory setup. Further, important data schemas will be presented which have been identified to make the data transfer between the laboratory devices and the computational analysis as coherent as possible.
Mark Nicholas Jones