Loading...

Design and Optimization of Microelectrodes Integrated on a Microchip for In Situ Acid Generation Using Machine Learning

Mohammadaghaie, Mehdi | 2025

0 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 58013 (08)
  4. University: Sharif University of Technology
  5. Department: Mechanical Engineering
  6. Advisor(s): Taghipoor, Mojtaba
  7. Abstract:
  8. The aim of this research is to design and optimize integrated microelectrodes on a chip for in-situ acid production and to investigate and utilize the capabilities of machine learning, particularly the reinforcement learning branch and genetic algorithms, in optimizing them with the goal of maximizing the number of electrodes on the chip, while ensuring that each operates independently. Initially, to validate and demonstrate the effectiveness of reinforcement learning using the DDQN algorithm, an agent was trained to optimize the geometric shape of an airfoil with the objective of maximizing the lift-to-drag ratio, showing that the agent can improve this ratio by over 100% in some cases. After proving the effectiveness of reinforcement learning in optimization, it is employed to optimize the geometric shape of microelectrodes for in-situ acid production, which is a suitable method for precisely altering PH using electric current and creating a redox reaction to produce H⁺ around the electrodes, independently for each electrode. The significance of this issue stems from the key role of PH, or the concentration of protons (H⁺), in many chemical and biological reactions as a catalyst, reactant, or product. For this purpose, two-dimensional and three-dimensional electrode configurations are simulated in COMSOL software, where the two-dimensional state is comb-shaped and the three-dimensional state consists of an anode in the center and a cathode ring around it. After creating the numerical model in COMSOL, it is coupled with Python, and using multiple numerical solutions and collecting labeled data, a deep neural network is trained to replace the numerical model. Subsequently, we will utilize the neural network and the PM-DQN algorithm to train the reinforcement learning agent. After completing the training, the agent will be used to approximately find the Pareto optimal front, and ultimately, the results will be compared with those of the genetic algorithm. We demonstrate that the agent can accurately estimate the Pareto front, achieving a score of 4.374 in the two-dimensional case compared to the genetic algorithm's score of 4.360 based on the hypervolume criterion, showcasing good performance. However, in the three-dimensional case, due to the complexity of the design space and the large number of possible actions, the agent's performance declines, and the genetic algorithm is preferred due to its lower computational burden
  9. Keywords:
  10. Machine Learning ; Reinforcement Learning ; Genetic Algorithm ; Deep Neural Networks ; Microelectrode Array ; Simulation ; Pareto Front ; Double Deep Q-Network (DDQN)Algorithm ; Pareto-Front-Based Multi-Objective Deep Q-Network (PM-DQN)Algorithm

 Digital Object List

 Bookmark

No TOC