Purpose

This study aims to introduce a vision-based model to generate droplets with auto-tuned parameters. The model can auto-adjust the inherent uncertainties and errors involved with the fabrication and operating parameters in microfluidic platform, attaining precise size and frequency of droplet generation.

Design/methodology/approach

The photolithography method is utilized to prepare the microfluidic devices used in this study, and various experiments are conducted at various flow-rate and viscosity ratios. Data for droplet shape is collected to train the artificial intelligence (AI) models.

Findings

Growth phase of droplets demonstrated a unique spring back effect in droplet size. The fully developed droplet sizes in the microchannel were modeled using least absolute shrinkage and selection operators (LASSO) regression model, Gaussian support vector machine (SVM), long short term memory (LSTM) and deep neural network models. Mean absolute percentage error (MAPE) of 0.05 and R² = 0.93 were obtained with a deep neural network model on untrained flow data. The shape parameters of the droplets are affected by several uncontrolled parameters. These parameters are instinctively captured in the model.

Originality/value

Experimental data set is generated for varying viscosity values and flow rates. The variation of flow rate of continuous phase is observed here instead of dispersed phase. An automated computation routine is developed to read the droplet shape parameters considering the transient growth phase of droplets. The droplet size data is used to build and compare various AI models for predicting droplet sizes. A predictive model is developed, which is ready for automated closed loop control of the droplet generation.

中文翻译：

---

基于人工智能的微流体系统液滴尺寸预测

 目的

本研究旨在引入一种基于视觉的模型来生成具有自动调整参数的液滴。该模型可以自动调整微流控平台制造和操作参数所涉及的固有不确定性和误差，获得精确的液滴生成尺寸和频率。

设计/方法论/途径

利用光刻方法制备本研究中使用的微流体装置，并在不同的流速和粘度比下进行了各种实验。收集液滴形状的数据来训练人工智能 (AI) 模型。

 发现

液滴的生长阶段表现出液滴尺寸的独特回弹效应。使用最小绝对收缩和选择算子 (LASSO) 回归模型、高斯支持向量机 (SVM)、长短期记忆 (LSTM) 和深度神经网络模型对微通道中完全发育的液滴尺寸进行建模。通过深度神经网络模型在未经训练的流数据上获得 0.05 的平均绝对百分比误差 (MAPE) 和R ² = 0.93。液滴的形状参数受到几个不受控制的参数的影响。这些参数被本能地捕获在模型中。

 原创性/价值

针对不同的粘度值和流速生成实验数据集。这里观察到的是连续相流速的变化，而不是分散相的流速变化。考虑到液滴的瞬态生长阶段，开发了自动计算例程来读取液滴形状参数。液滴尺寸数据用于构建和比较各种用于预测液滴尺寸的 AI 模型。开发了一个预测模型，可用于液滴生成的自动闭环控制。