Simulation and control of drug release on microneedle using machine learning technique

Designing a transdermal drug delivery (TDD) system is a difficult challenge in medical research. Transdermal methods provide the advantage of maintaining a steady drug concentration over time, which is important for drug control release applications such as vaccines and area-specific drugs. To advance this requirement, a microneedle (MN), a device capable of storing fluids on a micro-scale, is developed due to remarkable successes in minimally invasive, pain-free penetration through each skin layer. MN can be built into a system array in which multiple MNs are grouped. In the field, MNs are made up of various structures, drug delivery techniques, and manufacturing materials. In this thesis, we used, first, numerical simulation to solve the problem of drug release and MN structures, second, machine learning to accelerate numerical simulation and generate the optimal surrogate model for universal materials construction on MN simulations. The MN structures of solid, hollow, and semi-hollow MNs are simulated in the first part to observe these behaviors and construct a drug release control process. As a result, the solid structure can withstand steady pressure better than hollow or semi-hollow structures. On the other hand, hollow structure can release drugs more quickly than semi-hollow structure. Based on the simulation input, the results from hollow and semi-hollow structures could be utilized to characterize the drug-releasing control process. The model of control was demonstrated. The first part of this thesis provided impact research on the suitability of MN structures to each specific task and will be a crucial development on the future of drug release control mechanisms. In the second part of the thesis, the simulations of MN were created in both shapes and material variations. MN shapes such as pipe, rectangle, triangle, tapered-cone, beveled-tip, cone, and pyramid are simulated in 21 different materials. In each simulation, steady pressure was exerted on top of the MN as the post-processing result visualized the behavior of each shape and material while engaging the skin. In this part, we also emphasized the simulation's correctness by verifying the simulation with mesh independence analysis and comparing the results to previous literature with experimental results on MN. The MN injection process was validated and monitored force reaction during the insertion and relaxation periods. From these simulations, results with various materials were shown in order to identify which materials can withstand external pressure. Because numerical simulation requires excessive time effort and materials simulation is formalized by multiple material models, this part of the thesis attempts to accelerate the process and discover the optimal surrogate model that can compute all materials at once. Following that, the simulation results of only tapered-cone solid were extracted into a 3D structure for input in the machine learning (ML) development. ML's scope is defined as a regression problem for predicting the stress of the next time step based on prior time step data. Random Forest (RF), Gradient Boosting (GBoost), Artificial Neural Network (ANN), Graph Neural Network (GNN), Graph Attention Network (GAT), and PointNet model were used to test model development. During the training process, 1,024 data points are chosen in two ways: fixed-index points and random-index points. The results show that increasing the randomness of the training dataset has a considerable impact on the 3D-related structural problem. In each dataset, GBoost was the best model in fixed-index points, whereas GAT was the best model for a random-index dataset in testing with unseen material. Both data samples show the same patterns as ANN, GNN, PointNet, GAT, GBoost, and RF in consideration of prediction time. In conclusion, the graph model as GAT is recommended for future use in the development of more complicated neural network architecture and other material discovery purposes. This thesis has accomplished the study's objectives and presents additional applications in mathematical simulation and machine learning methodologies. The study results presented throughout the book would significantly influence and accelerate phenomena among micro-to-nano size devices, which could be applied in nano soft robotics or medical research applications.