Deep Learning Prediction of Surface Roughness in Multi‐Stage Microneedle Fabrication: A Long Short‐Term Memory‐Recurrent Neural Network Approach

Abstract

Microneedles (MNs) have emerged as a promising tool for transdermal drug delivery, with surface roughness being a critical parameter affecting coating adhesion and drug delivery efficiency. This study aimed to predict and optimize the surface roughness of MNs fabricated through a multi‐stage process (MFP) using a long short‐term memory (LSTM)‐based recurrent neural network model. The fabrication process involved three stages: 3D printing of the master mold, creation of a polydimethylsiloxane  (PDMS) female mold, and fabrication of polyvinyl alcohol (PVA) MNs. Roughness metrics were extracted from microscopy images at each stage, and the LSTM model leveraged sequential dependencies to predict the progression of roughness from the master mold ( R 1 ) to the PDMS mold ( R 2 ) and the final PVA MNs ( R 3 ). The model was trained and validated on 50 samples, achieving a R 2 score of 0.54 on the training set and 0.64 on the test set for predicting R 3 . Roughness categorization accuracy reached 70% on both the training set and the test set. Sensitivity analysis identified height as the most influential parameter, contributing to an average impact of 7.88% on R 3 predictions, followed by base diameter at 2.53%, R 2 at 1.37%, and R 1 at 1.01%.