Transfer Learning-Based Microfluidic Design System

for Concentration Generation

 Author
• Weiqing Ji, Department of Computer Science and Technology, Tsinghua University

• Tsung-Yi Ho, Department of Computer Science, National Tsing Hua University

• Hailong Yao, Beijing National Research Center for Information Science and Technology, Tsinghua
University
 Outline
• Introduction
• Preliminary And Problem Formulation
• Overview Of The Proposed Method
• Methods
• Computational Simulation Results
• Conclusion
 Introduction
• Originally, solutions of custom gradients were prepared by manual pipetting, which is
time-consuming and error-prone.
• More effective concentration generation methods :
• liquid handling robots
• In droplet-based microfluidic biochips, droplet splitting and merging are electrically
controlled to obtain the required concentration solution.
• In multi-layer flow-based microfluidic biochips, the mixing process for specific
concentration solutions is controlled by switching the valves between control and
fluidic layers.
Introduction
 Introduction
• The custom concentrations can be obtained by adjusting both the biochip’s internal
structure and the length of outlets.
• An artificial neural network (ANN)-based method:
• To design specific biochips automatically, which accurately generates the custom
concentration.
• Used to adjust the outlet length of the biochip to cover different concentration
requirements.
• Major contributions :
• The first transfer learning method for custom concentration generation.
• The fine-tune strategy is evaluated and adopted.
• Using the proposed transfer learning method to generating a database.
 Preliminary and Problem Formulation
• Notations “Domain” and “task” are often used to define and classify transfer learning
methods.
• Domain D : a feature space X and a marginal probability distribution P(X).
• Task T : a label space Y and an objective predictive function f ().

• Transfer learning aims to improve the learning process of the target predictive function f T()
in DT using the knowledge in DS and TS.
• Transfer learning approaches can be categorized into several cases :
• Model adaptation
• Heterogeneous transfer learning
• Model transfer
• Multi-task learning
 Preliminary and Problem Formulation
• A deep learning model is trained using the dataset in the source task.
• The trained model is called the pre-trained network.
• Weights of the pre-trained network consist of the learned features, which can be shared by
similar tasks.
• Transfer learning is adopted for deep learning :
• The freeze and train strategy.
• The fine-tune strategy
• adopted in our proposed method.
 Preliminary and Problem Formulation
• The ANN-based design method can significantly enhance accuracy of concentration
generation.
• When the sample and reagents changes, or the number of inlets/outlets need to be
modified, the whole system needs to be reconstructed.
• The deep-learning-based design method for microfluidic biochips :
• Input: Expected custom output concentration values at outlets.
• Output: The DXF [20] format of the designed biochip, which can be directly used for
fabrication.
• Objective: Minimize the errors between the actual and expected concentration values.
• Constraint: The minimum channel width and spacing design rules need to be satisfied.
Overview of the Proposed Method
 Method
• Problem Analysis:
• The transfer learning method can transfer the knowledge learned from the task of one
microfluidic biochip (the source task) to the tasks of the other microfluidic biochips (the target
tasks), which reduces the amount of training data required by the target tasks.

• Dataset Generation:
• Let C1, C2, and C3 denote the output concentrations at the three outlets.
• We set an evenly distributed concentration requirements as C1,C2,C3 = 0.07+0.15k
∈ [0,1], C1>= C2>= C3, k ∈ N, which in total are 84 requirements.
• These concentrations are used to query the database in [10] for obtaining the initial
biochips.
• One microfluidic biochip is randomly selected from 84 biochips as the source task.
• The remaining 83 microfluidic biochips are used as target tasks.
 Method
• Database Generation:

• Query Procedure:
 Computation Simulation Results

• When the size of dataset is identical, the accuracy rate of the fine-tune strategy is higher
than that of the freeze strategy.
• As the runtime for sample generation is 128 days in [11], our proposed method reduces
the runtime by 96%.
• The mean error calculated by our proposed method is much smaller than that of [10],
with an average reduction rate of 83%.
Computation Simulation Results
Computation Simulation Results