Can machine learning recover selectivity when the electrochemical sensor chemistry is not perfectly selective?

The electrochemical problem

Electrochemical sensors for PFAS often rely on molecularly imprinted polymer concepts, but actual selectivity can be imperfect. In this study, electropolymerized ortho-phenylenediamine sensors showed template-dependent sensitivity and cross-reactivity among PFOS, PFOA, PFBS, and PFBA.

The resulting question is practical: if the chemistry creates reproducible but overlapping response patterns, can machine learning extract analyte identity from the full electrochemical response even when no single sensor is chemically specific enough?

What the researchers did

The authors measured PFAS-dependent responses by differential pulse voltammetry using a ferrocenemethanol redox probe. They examined the underlying sensor selectivity, including the observation that some imprinted sensors responded to related PFAS rather than only the template molecule.

The DPV data were then used to train a machine-learning classification model to distinguish four PFAS species.

What they found

The classifier reached 95% accuracy in distinguishing PFOS, PFOA, PFBS, and PFBA. The paper therefore demonstrates a useful distinction between chemical selectivity at the material interface and system-level selectivity obtained from the combined sensor-plus-model response.

The analytical platform also reported low detection limits for the individual PFAS targets, while the classification result addresses the separate problem of telling similar analytes apart.

Why this matters

Electrochemical sensing is often limited by overlapping responses rather than by an absence of signal. Machine learning can be valuable when the sensor produces stable multidimensional fingerprints that are individually ambiguous but collectively informative.

This suggests a design philosophy in which sensing materials and algorithms are co-developed. The goal need not always be a perfectly selective electrode if a reproducible sensor array or response pattern can support reliable discrimination.

One important caution

A 95% classification result is not equivalent to universal PFAS identification. Performance depends on the analytes, concentration ranges, matrix, sensor fabrication, and training distribution represented in the study.

Environmental samples can contain many interferents outside the four-class problem. Extending system-level selectivity to open-world monitoring therefore requires validation under broader and more variable sample conditions.