Dye-sensitized solar cells under ambient light powering machine learning: towards autonomous smart sensors for the internet of things

• Published in: Chemical science (Cambridge) Vol. 11; no. 11; pp. 2895 - 296
• Main Authors: Michaels, Hannes, Rinderle, Michael, Freitag, Richard, Benesperi, Iacopo, Edvinsson, Tomas, Socher, Richard, Gagliardi, Alessio, Freitag, Marina
• Format: Journal Article
• Language: English
• Published: England Royal Society of Chemistry 13.02.2020; The Royal Society of Chemistry
• Subjects: Artificial intelligence; Artificial neural networks; Chemistry; Dye-sensitized solar cells; Dyes; Electrolytic cells; Energy conversion efficiency; Energy sources; Fluorescent lamps; Harvesters; Illumination; Image acquisition; Image classification; Inference; Internet of Things; Light; Machine learning; Microcontrollers; Neural networks; Photons; Photovoltaic cells; Portable equipment; Power management; Smart sensors
• ISSN: 2041-6520; 2041-6539; 2041-6539
• DOI: 10.1039/c9sc06145b

The field of photovoltaics gives the opportunity to make our buildings ''smart'' and our portable devices "independent", provided effective energy sources can be developed for use in ambient indoor conditions. To address this important issue, ambient light photovoltaic cells were developed to power autonomous Internet of Things (IoT) devices, capable of machine learning, allowing the on-device implementation of artificial intelligence. Through a novel co-sensitization strategy, we tailored dye-sensitized photovoltaic cells based on a copper( ii / i ) electrolyte for the generation of power under ambient lighting with an unprecedented conversion efficiency (34%, 103 μW cm −2 at 1000 lux; 32.7%, 50 μW cm −2 at 500 lux and 31.4%, 19 μW cm −2 at 200 lux from a fluorescent lamp). A small array of DSCs with a joint active area of 16 cm 2 was then used to power machine learning on wireless nodes. The collection of 0.947 mJ or 2.72 × 10 15 photons is needed to compute one inference of a pre-trained artificial neural network for MNIST image classification in the employed set up. The inference accuracy of the network exceeded 90% for standard test images and 80% using camera-acquired printed MNIST-digits. Quantization of the neural network significantly reduced memory requirements with a less than 0.1% loss in accuracy compared to a full-precision network, making machine learning inferences on low-power microcontrollers possible. 152 J or 4.41 × 10 20 photons required for training and verification of an artificial neural network were harvested with 64 cm 2 photovoltaic area in less than 24 hours under 1000 lux illumination. Ambient light harvesters provide a new generation of self-powered and "smart" IoT devices powered through an energy source that is largely untapped. Indoor light harvesters enable machine learning on fully autonomous IoT devices at 2.72 × 10 15 photons per inference.