Implementação de rede neural na caracterização de IPMC´s com líquidos iônicos de alquilimidazólio incorporados: estudo de viabilidade de ferramenta auxiliar

Abstract

The present study aims to implement and evaluate the feasibility of an artificial neural network algorithm in Python to classify IPMC samples. For this work it was considered Nafion-based IPMC samples with platinum electrodes. One of the IPMC’s applications is its use as soft actuators and sensors due to its smooth movements similar to biological muscles. This bending mechanism occurs because of the cationic diffusion inside Náfion®’s ionomeric channels dragging solvent molecules causing a difference in the solvent concentration along the sample thickness [12, 13, 14]. Based on studies and electromechanical characterizations, it was observed that the hydration level and external conditions influence in the actuator performance [14]. The loss of water to the environment during the actuation cycles implies a drop in the IPMC effectiveness which is a challenge for commercial applications. The incorporation of the so-called ionic liquids, in this case was used the 1-butyl-3- ethylimidazole chloride (BMIM.Cl), may be one possible solution so the IPMC works with its higher capacity for longer periods. Besides the electromechanical characterizations in which the incorporation of BMIM.Cl proved to be very promising, dynamic mechanical oscillatory shear measurements with torsional rectangular geometry were also carried out in order to study the IPMC’s viscoelastic response. Therefore, the storage modulus (G’) as a function of frequency (w), previously obtained in torsion tests performed in ARES rheometer from TA Instruments, were used as input data in the neural network. The methodology used was the adaptation of the of the publicly available neural network code in the Machine Learning tutorial for Python programming [28]. The neural network has three layers, two hidden and one output layer, densely connected. Thus, after the training and validation of the artificial neural network, the algorithm gives us the accuracy. Firstly, using the relu and sigmoid activation functions, the accuracy measured was between 16% and 19%. After the study of the influence of the activation functions and its change in the hidden layers and in the output layer it was obtained a accuracy higher than 95%. Furthermore, random noisy data were generated in order to simulate different analysis of the same sample containing a variability. This simulated was used to evaluate the accuracy with the increase of the noise intensity. The algorithm showed that the accuracy in the capacity of rightly classify the IPMC samples drops with higher intensities noises. So, although the problem herein approached was very simple and the quantity of data available was relatively small, from preliminary results, the implementation ofthe neural network algorithm showed to be a viable tool with high accuracy level to assist in the characterization of IPMC samples as long as the noise range in relation to the experimental data is within a certain limit. It is of great relevance to emphasize the importance of a significant amount of data to get more conclusive results.