Jul 21, 2011
As-fabricated nanosensor learns to classify solvents
Sensors comprised of nanoscale building blocks are studied by many groups who aim to exploit their huge specific surface area to achieve remarkable sensitivity to the environment. The most common goal is to optimize the sensitivity of the sensor for a particular task such as detecting toxic gases or specific biological species. Arguably, one thing that is common to all nanosystems, especially those fabricated through self- or directed-assembly processes, is that no two nanoscale sensors are identical and for the realistic proliferation of nanosensors, tolerance to such variance and moreover, to defects is essential. Perhaps there is a role for some mathematics here?
One approach explored by researchers at the Nokia Research Center in Cambridge, UK, is to employ silicon nanowire sensors together with a statistical machine-learning technique that is more commonly associated with pattern recognition than in nanotechnology. The Nokia Research team fabricated an array of nanowire sensors and constructed a simple table-top demonstrator that was able, after appropriate training of the system, to correctly discriminate between common laboratory solvents with greater than 95% accuracy.
What is striking about this approach is that the sensors were used as-fabricated. They were not chemically functionalized, they were not optimized for the task, and in some cases the arrays even contained defective sensors, but machine learning can be used to discover how the array behaves collectively in response to the chemicals, and usefully this information can then be used to extrapolate that learning to classify in real-time.
Extracting value from disorder
Learning from the collective behaviour of an array of nanoscale elements, whether it be for sensing, computing or some other task, in response to a particular input or combination of inputs could be a powerful technique for nanosystems in general. If nanotechnologies are to take the giant leap towards widespread commercialization then low-cost manufacturing techniques coupled with defect-tolerant computation based on machine learning may be the way to go.
The researchers presented their results in the journal Nanotechnology.
About the author
The work was conducted by Drs AO Niskanen, A Colli, R White, HW Li, E Spigone and JM Kivioja at Nokia Research Center, Cambridge, UK.