Machine Learning Techniques for Chemical Identification Using Cyclic Square Wave Voltammetry.

Area of Science:

• Electrochemistry
• Analytical Chemistry
• Machine Learning

Background:

• Cyclic square wave voltammetry (CSWV) is a versatile electroanalytical technique for detecting diverse compounds.
• A major limitation for CSWV is the lack of robust, generalizable classification algorithms for sample identification.
• Existing classification methods for CSWV data often lack sufficient performance.

Purpose of the Study:

• To develop and evaluate advanced machine learning and deep learning models for classifying CSWV data.
• To enhance the accuracy and generalizability of chemical identification using electroanalytical signals.
• To create an automated workflow for processing CSWV data and training predictive models.

Main Methods:

• Development of Long Short-Term Memory (LSTM) and Fully Convolutional Networks (FCNs) for voltammogram classification.
• Training and testing models on various CSWV datasets.
• Comparative analysis against previously used algorithms.
• Utilizing class activation maps to interpret model decisions.
• Validation on new, unseen experimental data.

Main Results:

• LSTM and FCN models demonstrated superior performance compared to existing algorithms.
• Achieved area under the curve (AUC) values greater than 0.99 in receiver operating characteristic (ROC) analyses for several datasets.
• Models successfully generalized to new experimental data.
• Class activation maps provided insights into the networks' decision-making processes.

Conclusions:

• Machine and deep learning, specifically LSTM and FCNs, provide a high-performance solution for CSWV data classification.
• These models significantly enhance the sensitivity and specificity of chemical identification.
• The developed automated method and tools support field-deployable compound identification using CSWV.