A collaborative team from NIMS and Tokyo University of Science has achieved a remarkable feat in artificial intelligence by developing a device that mimics brain-like information processing using a technique called few-molecule reservoir computing. This groundbreaking innovation harnesses the molecular vibrations of a select group of organic molecules.
Utilizing this device for predicting blood glucose levels in diabetic patients has yielded exceptional results, surpassing the accuracy of existing AI devices. Their findings are documented in the journal Science Advances.
As machine learning applications continue to expand across industries, there’s a growing demand for AI devices that are not only powerful but also energy-efficient and compact. To address this need, researchers have turned to physical reservoir computing, leveraging material and device properties for neural information processing. However, a key challenge has been the large size of existing materials and devices.
The research team has achieved a breakthrough by implementing physical reservoir computing based on surface-enhanced Raman scattering. This technique utilizes the molecular vibrations of a few organic molecules, such as p-mercaptobenzoic acid (pMBA). Information input is facilitated through ion gating, which controls hydrogen ion adsorption onto pMBA molecules via voltage application.
The variations in molecular vibrations of pMBA molecules, influenced by hydrogen ion adsorption, enable memory storage and nonlinear waveform transformation for computation. Remarkably, this sparse assembly of pMBA molecules has learned patterns in blood glucose level changes from a diabetic patient’s data spanning approximately 20 hours. It then accurately predicted subsequent fluctuations over the next five minutes with a 50% error reduction compared to existing devices.
This study showcases the remarkable computational capabilities achievable with minimal organic molecules, rivaling traditional computers. This technological advancement holds significant practical implications, paving the way for low-power AI terminal devices that can integrate with various sensors for widespread industrial applications.