What and how can machine learning help to decipher mechanisms in molecular electrochemistry?

Maximizing Electrochemical Information: A Perspective on Background-Inclusive Fast Voltammetry

“…To lend additional credence to the idea of forgoing background subtraction, we point to studies in the mechanistic electrochemistry field. As opposed to using background-subtracted voltammograms to train machine learning models to predict analyte identity and concentration, fundamental electrochemistry studies use background-inclusive voltammograms to fit simulated and experimental data, including nonfaradaic current. − These reports further demonstrate the utility of nonfaradaic information in models of electrochemical processes beyond concentration quantification. For example, areas of voltammograms not typically used in the manual assignment of electrochemical reaction mechanisms are now being used by deep learning classifiers for automated mechanistic assignment .…”

“…Based on the publications reviewed here on the importance of nonfaradaic information and the versatility of waveforms (sweeps and pulses) in voltammetry, the next major advances for in vivo voltammetry appear likely to come from background-inclusive approaches paired with machine learning. There are many recent examples of movement in this direction inside and outside of the chemical neuroscience community. ,,,,− ,,,,,, …”

“…There are many recent examples of movement in this direction inside and outside of the chemical neuroscience community. 11,14,16,20,[44][45][46]64,65,71,72,74,79 If there is a solution to the pervasive problems that have plagued voltammetry for decades preventing the full electrochemical exploration of the chemical communication systems of the brain and beyond, recent evidence points to a need to reconsider the use of background subtraction. Broadly speaking, all practitioners of voltammetry should consider maximizing the information inherent in their experimental data and complementing domain knowledge with their analysis toolkit of choice.…”

“…The most advanced models today consist of far more robust experimental designs with training sets containing multiple concentrations of analytes, metabolites, H + and other ions, multiple electrodes, and so on, across thousands of voltammograms. , As state-of-the-art (i.e., deep learning) models are developed, ,,, electrochemists will also likely find greater success in maximizing the information content of data acquisition. Examples include the fusing of multiple data sources, the ability to perform inference on out-of-distribution data, and the use of physics-informed and probabilistic models. These areas are likely to yield complementary advances for machine learning and voltammetry that extend beyond neurochemical detection toward electroanalytical chemistry writ large .…”

“… 38 , 48 As state-of-the-art (i.e., deep learning) models are developed, 35 , 37 , 39 , 46 electrochemists will also likely find greater success in maximizing the information content of data acquisition. Examples include the fusing of multiple data sources, 45 the ability to perform inference on out-of-distribution data, 38 and the use of physics-informed 43 and probabilistic 49 models. These areas are likely to yield complementary advances for machine learning and voltammetry that extend beyond neurochemical detection toward electroanalytical chemistry writ large .…”