A single-fibre computer enables textile networks and distributed inference

• 成果类型: Article
• 署名作者: Gupta, Nikhil; Cheung, Henry; Payra, Syamantak; Loke, Gabriel; Li, Jenny; Zhao, Yongyi; Balachander, Latika; Son, Ella; Li, Vivian; Kravitz, Samuel; Lohawala, Sehar; Joannopoulos, John; Fink, Yoel
• 署名单位: Massachusetts Institute of Technology (MIT); Massachusetts Institute of Technology (MIT); Massachusetts Institute of Technology (MIT); Stanford University; Rhode Island School of Design; Brown University; Massachusetts Institute of Technology (MIT); Massachusetts Institute of Technology (MIT); Massachusetts Institute of Technology (MIT)
• 刊物名称: Nature
• ISSN/ISSBN: 0028-1279
• DOI: 10.1038/s41586-024-08568-6
• 发表日期: 2025-03-06
• 关键词: fusion

摘要

Despite advancements in wearable technologies1,2, barriers remain in achieving distributed computation located persistently on the human body. Here a textile fibre computer that monolithically combines analogue sensing, digital memory, processing and communication in a mass of less than 5 g is presented. Enabled by a foldable interposer, the two-dimensional pad architectures of microdevices were mapped to three-dimensional cylindrical layouts conforming to fibre geometry. Through connection with helical copper microwires, eight microdevices were thermally drawn into a machine-washable elastic fibre capable of more than 60% stretch. This programmable fibre, which incorporates a 32-bit floating-point microcontroller, independently performs edge computing tasks even when braided, woven, knitted or seam-sewn into garments. The universality of the assembly process allows for the integration of additional functions with simple modifications, including a rechargeable fibre power source that operates the computer for nearly 6 h. Finally, we surmount the perennial limitation of rigid interconnects by implementing two wireless communication schemes involving woven optical links and seam-inserted radio-frequency communications. To demonstrate its utility, we show that garments equipped with four fibre computers, one per limb, operating individually trained neural networks achieve, on average, 67% accuracy in classifying physical activity. However, when networked, inference accuracy increases to 95% using simple weighted voting.