High-precision blood glucose level prediction achieved by few-molecule reservoir computing

With the expansion of machine learning applications in various industries, there's an escalating demand for AI devices that are not only highly computational but also feature low-power consumption and miniaturization. Research has shifted towards physical reservoir computing, leveraging physical phenomena presented by materials and devices for neural information processing. One challenge that remains is the relatively large size of the existing materials and devices.

Our research has pioneered the world's first implementation of physical reservoir computing that operates on the principle of surface-enhanced Raman scattering, harnessing the molecular vibrations of merely a few organic molecules. The information is inputted through ion-gating, which modulates the adsorption of hydrogen ions onto organic molecules (p-mercaptobenzoic acid, pMBA) by applying voltage. The changes in molecular vibrations of the pMBA molecules, which vary with hydrogen ion adsorption, serve the function of memory and nonlinear waveform transformation for calculation. This process, using a sparse assembly of pMBA molecules, has learned approximately 20 hours of a diabetic patient's blood glucose level changes and managed to predict subsequent fluctuations over the next 5 minutes with an error reduction of about 50% compared to the highest accuracy achieved by similar devices to date.

The outcome of this study indicates that a minimal quantity of organic molecules can effectively perform computations comparable to a computer. This technological breakthrough of conducting sophisticated information processing with minimal materials and in tiny spaces presents substantial practical benefits. It paves the way for the creation of low-power AI terminal devices that can be integrated with a variety of sensors, opening avenues for broad industrial use.