DLDNN is a design automation platform for deterministic lateral displacement (DLD) devices that incorporates the power of neural networks and multi-objective genetic algorithm.
The design automation algorithms convert the user-defined specification to geometry properties and flow rate to readily meet the desired criteria. To boost the function evaluation speed, one can utilize neural networks as the surrogate model instead of the whole physics/experimentation. This alternative facilitates the design process of a DLD device to reach the best possible configuration, especially when combined with an evolutionary algorithm. As presented in Figure 2, design automation inputs are the diameters of particles to be separated (D1 and D2), desired constraints to be applied on f, N, Re (Cf , CN, and CRe), and finally the trade-off between flexibility and stability of the design (Φ). Flexibility indicates the ability of the proposed design to cover a wide range of critical diameters while stability specifies its immunity to change due to minor flow rate fluctuations. These two terms were included to consider the possible device behavior after fabrication since it can be altered by adjusting the Reynolds number. After defining the inputs, a multi-objective optimization algorithm coupled with a pre-trained FCNN tunes the f, N, Re, and G parameters to reach the optimized design. In this design automation tool, an NSGA3 with five reference directions and population size of 260 was chosen and implemented in the Pymoo library in Python 3. The pre-trained FCNN was used for direct critical diameter prediction (direct neural network) since extracting critical diameter from CNN outputs slows down the optimization process. Finally, the tuned parameters in addition to bandwidth (BW) are extracted and presented in the outputs. BW is the difference between the maximum and the minimum critical diameter for that design serving as an index of flexibility. Note that the parameter G is used in optimization to nondimensionalize the critical diameter matching the developed neural network’s Dc. Furthermore, the pre-trained CNN was attached to the end of the automation process to provide the ability to simulate the particle trajectory in the optimum design for a given number of periods. This feature helps visualize the particle behavior at the full length of the device.
More information about this method can be found in our DLDNN paper.
- DLD_utils.py: Containing the function for extracting data from numerical simulation and mapping the generated field.
- DLD_env.py: Containing the function for post-processing of the data from numerical simulation.
- particle_trajectory.py: Simulate particle trajectory from the velocity fields in horizontal and vertical directions.
- generate_data.py: Dataset generation for the convolutional neural network.
- Conv_base.py: The convolutional neural network class containing the network architecture.
- Conv_net_train.py: The training process of the convolutional neural network.
- Temp9: The trained CNN model weights and bias.
- Direct_network_generate_data.py: Changes the labels to fields data-set into fields to critical diameters data-set.
- Direct_NN.py: The Fully-connected neural network class containing the structure of the neural network, training process.
- DNN_model_hlayers8_nodes_128.h5: The trained Direct neural network weights and bias.
- Design_Aut.py: Main code of the design automation platform.
- If it is needed we can provide the data sets that were used in this study.