This post was originally published on Nature.com.
The story behind our paper; “A hardwired machine learning processing engine fabricated with submicron metal-oxide thin-film transistors on a flexible substrate”, started with a question of how to make billions of everyday things smart. For example, food packaging, deodorants, detergents, bandages, – the list goes on. The simplest response to this question was to attach some cheap electronics (~$5) to them to make them smarter. But, the definition of low-cost electronics is relative: adding $5 to a $200 handbag will not make much of a difference in the selling price but adding the same electronics to a 50-cent milk bottle is a showstopper.
Despite the great success of conventional electronics, the high costs of Si-based semiconductor fabrication make it impossible to manufacture them at the cost point necessary to make ultra-cost-sensitive things smarter. This is where flexible (also known as organic) electronics comes to the rescue. Unlike conventional electronics, flexible electronics are inherently thin, and conformable. Most importantly, the manufacturing system has also been designed to enable ultra-low-cost production with rapid turnaround times. In this way, it is possible to design the most optimized processor for the job, rather than a general processor.
From a technology point of view, to make a “dumb” thing smarter, three modalities are essential: sensing, processing, and communicating. Printed sensors have been around for a long time for communication purposes, and flexible NFC antennas and circuitry are being developed both in academia and industry. What is missing is a flexible processor.
So, we embarked on the fully custom design, implementation, and fabrication of a microprocessor with thin-film transistors (TFTs) on a flexible substrate. We define and use the term “Natively Flexible Processing Engine (NFPEs)” to describe this processor throughout the Article.
Fig. 1 Micrograph of the ML Natively Flexible Processing Engine (NFPE) as a flexible IC
In the Article, we demonstrate our NFPE can be directly attached to an e-nose sensor to process the data and make decisions based on it. It is a machine learning (ML) processing engine that reads the e-nose sensor data for odor recognition application and recognizes the odor intensity level. The ML NFPE is the implementation of a resource-efficient ML algorithm that we also developed; ‘Univariate Bayes feature voting classifier’. We fabricated and validated the ML NFPE as a flexible IC using PragmatIC’s 0.8µm process with n-type metal-oxide TFTs, as shown in Fig. 1. The ML NFPE is the most complex digital circuit fabricated with metal-oxide TFTs to date.
Fig. 2 Various ML NFPEs can be fabricated on polyimide substrate for various applications, each of which have their own requirements and learned parameters.
Although the ML NFPE was developed for a specific odour-based application, the methodology is generic enough to be adapted to other odour-based applications (for example, food packaging, wound dressing, room air quality detection and so on) to develop a new ML NFPE specific to the application under consideration (see Fig. 2).
Read the full paper to find out more about the technology, and how it could help bring intelligence to millions of everyday objects.
Read the Full Paper