Fig. 6
From: Machine learning based multi-parameter droplet optimisation model study
GP-BO optimisation flow chart.(A) Initialisation of the control parameters is performed to normalise the parameter values to between [0–1]. Shown above and below are scatter plots of the initial control parameter distributions for the millimetre-scale inkjet device as well as the microfluidic device, respectively. (B) 140 droplet images generated under different parameter controls for each of the millimetre-scale inkjet device as well as the microfluidic device were used as the dataset. The droplet images corresponding to the initialised control parameter values are input into the computer vision module, and the computer vision uses the watershed segmentation technique to separate the droplets from the background in the droplet images to obtain all the droplet pixels in the images. (C) The separated droplets are quantified by the objective function, and the separated droplets are scored for geometric roundness, yield, and size uniformity detection and input into the BO module. (D) The control parameters are modelled using a GP, and the fit of the GP to the objective function is first examined, and when the fit is perfect, the input objective function values and control parameter values are BO’d and the predicted parameter values with the smallest objective function values, which consist of the loss of roundness, the loss of yield, and the loss of uniformity of size, are output.
