Machine Learning Optimizes Polymer Analysis

By Soham NandiReviewed by Susha Cheriyedath, M.Sc.Aug 30 2024

In an article published in the journal Electrochimica Acta, researchers presented a novel approach that integrated a convolutional neural network (CNN) model with broadband dielectric spectroscopy (BDS) to predict the electrical equivalent circuit (EEC) topology of polymer membranes, particularly ion exchange membranes (IEM).

This method aimed to reduce user bias, enhance data analysis, and improve the design and characterization of polymers used in renewable energy conversion technologies, fuel cells, and energy storage applications.

Background

Polymer membranes, particularly IEMs, play a critical role in various renewable energy conversion technologies, such as fuel cells and photoelectrochemical cells, by enabling efficient ion transport and energy conversion processes. Understanding the dynamic behavior of these polymers is essential for optimizing their performance.

BDS is a widely used technique to study the electrical response of polymers, providing valuable insights into their molecular dynamics and relaxation processes. However, the traditional method of determining the EEC for interpreting BDS data is often subjective, relying heavily on the experimenter's expertise and biases. This subjectivity can lead to inaccuracies in modeling complex polymer systems, as the correct EEC topology might be difficult to identify due to overlapping relaxation processes or insufficient experimental data.

To address these challenges, this paper introduced a novel approach that employed machine learning to predict the EEC topology, thereby reducing bias and improving the accuracy of BDS data analysis. This method aided experts in extracting more reliable information from their data and made BDS more accessible to less experienced researchers.

Membrane Preparation and Impedance Measurement Setup

The researchers investigated two ion exchange membranes, namely, Nafion® N115 (cationic) and Fumasep® FAA-3-PK-75 (anionic). Both membranes were cleaned, activated, and stored using specific procedures. Nafion® N115 was treated with hydrogen peroxide and sulfuric acid, while Fumasep® FAA-3-PK-75 underwent hydroxide pumping with sodium hydroxide.

The membranes were then subjected to a four-terminal sensing (4T) technique, also known as Kelvin sensing, for electrical impedance measurement. This method, which separated current and voltage electrodes, allowed precise measurements by eliminating lead and contact resistance. The experiments were conducted at room temperature and controlled humidity, utilizing a potentiostat and a custom setup to ensure accurate readings of the membrane's impedance.

Equivalent Electrical Circuit Determination

The authors presented a machine learning approach to determine the EEC for a given impedance dataset, employing a two-step methodology. First, a neural network was trained using supervised learning to predict the circuit topology.

The training data consisted of impedance data paired with circuit descriptions, focusing on topologies generated through random sampling of candidate circuits derived from a universal circuit model. The model outputted binary vectors representing the presence or absence of specific circuit elements, with the network architecture utilizing a two-dimensional CNN that processed multi-scale features of impedance data.

In the second step, a global optimization method, specifically the trust region approach, was applied to determine the parameters of the predicted circuit. The optimization was constrained within practical bounds, ensuring the parameters remained within feasible ranges. To avoid local minima in this non-convex problem, the method was applied multiple times from different random starting points (multi-start method), selecting the best result based on the lowest fitting error.

This combined approach effectively predicted the topology and parameters of EECs from impedance data, providing accurate models for electrical circuits commonly encountered in electrochemical impedance spectroscopy (EIS).

Performance Analysis and Model Validation

The performance of the neural network model was assessed in identifying circuit topologies and equivalent circuits. The dataset was split into training, validation, and test sets, with the neural network trained using the Adam optimizer for around 300 epochs.

The model's classification accuracy for top-1 predictions was about 40%, primarily due to non-identifiability issues where different topologies produce similar impedance spectra. However, top-5 accuracy increased to approximately 80%, indicating that the correct topology was among the top five predictions.

The model was benchmarked against other machine learning models and outperformed them in top-1 prediction accuracy across various circuit classification tasks. For regression results, the mean absolute percentage error (MAPE) was used to evaluate the fitting accuracy of the predicted circuits.

Despite the low top-1 classification accuracy, the fitting error for the top-3 predictions was comparable to the true circuit, suggesting that the low classification accuracy was not due to poor model performance but to the inherent mathematical equivalence of different topologies.

Validation was performed using experimental data from two polymer membranes, Nafion® 115 and Fumasep® FAA-3-PK-75. The neural network's predictions were validated against this data, and the resulting parameter sets for the EECs showed good agreement with the measured impedance spectra, with a coefficient of determination (R²) close to 0.998 and 0.999, respectively.

Conclusion

In conclusion, the researchers presented a groundbreaking approach by integrating a CNN with BDS to predict the EEC topology of polymer membranes. This method enhanced the accuracy and efficiency of analyzing polymer behavior, particularly in renewable energy applications, by reducing user bias and improving data interpretation.

The model’s ability to accurately identify circuit topologies and parameters demonstrated its potential to revolutionize polymer membrane research, offering a valuable tool for both experts and novices in the field of renewable energy and material science.