Study of sensors using machine learning
No Thumbnail Available
Authors
Meeting name
Sponsors
Date
Journal Title
Format
Thesis
Subject
Abstract
[EMBARGOED UNTIL 12/01/2026] Effective long-term monitoring of nutrient pollutants such as nitrate and phosphate remains a central challenge in environmental and agricultural management. Excess concentrations of these anions in waterways contribute to eutrophication, soil degradation, and significant economic losses associated with impaired ecosystem services. While ion-selective electrodes (ISEs) offer a promising route for decentralized and continuous sensing, the dependence on highly selective membrane materials and the susceptibility of these membranes to drift, fouling, and degradation limit the scalability and long-term stability of traditional ISE-based sensors. To address this gap, this study investigates the feasibility of multi-ion quantification using pulsed-current sensing approach coupled with machine learning, thereby reducing the reliance on membrane specificity. In this work, three nonspecific ion-responsive electrodes were fabricated using membrane support materials that are widely available and mechanically robust – poly(vinyl acetate) (PVA), poly(vinyl chloride) (PVC), and poly(acrylonitrile) (PAN). Each electrode was subjected to potential-time measurements under pulsed-current excitation in solutions containing nitrate, phosphate, and chloride ions. These individual ion response profiles formed the basis profiles for constructing synthetic dataset representing mixtures of the target ions, enabling controlled exploration of multi-ion interactions without the experimental complexity of preparing large numbers of mixed standards. To evaluate the capacity of nonspecific sensors to resolve multi-ion signals, a suite of machine learning algorithms–including tree-based regressors and fully connected neural networks–was applied to the combined electrode responses. Model performance was assessed in terms of predictive accuracy, robustness to noise, and sensitivity to electrode-specific variations. Initial results demonstrate that machine learning can extract meaningful compositional information from pulsed-current responses, providing a proof of concept for low-selectivity, multi-ion sensing without the need for specialized membrane chemistries. However, several limitations remain, particularly the variability in electrode response arising from differences in membrane morphology, the constrained diversity of the training dataset, and idealized nature of synthetic mixture generation. Overall, this study highlights both the promise and present challenges of integrating pulsedcurrent electrochemical sensing with machine learning for environmental monitoring applications. The findings point toward future directions involving improved electrode fabrication reproducibility, the collection of larger and more representative training datasets, and the transition from synthetic data to experimentally validated mixture responses. Collectively, these advancements will be crucial for the realization of practical, durable, and cost-effective multi-ion monitoring systems suitable for field deployment.
Table of Contents
PubMed ID
Degree
M.S.