Nanomagnetic arrays for novel computation

我们已经开发了一种使用廉价、低功率(~2 mW)激光器的全光纳米磁性写入技术。与其他光磁方法不同，它不需要磁场、奇异的材料，也不需要强大的激光--这使得它便宜、对设备友好、可工业扩展。同时，我们开发了世界上第一个纳米磁阵列神经形态计算方案，提供低功率神经网络功能2。这种概念验证效果很好，但神经网络权重更新依赖于缓慢变化的磁场，并且缺乏单纳米磁铁的分辨率。光学写入和读取现在允许我们直接写入网络权重(纳米磁铁状态)。这从根本上提高了运营速度并解决了设备兼容性问题，具有巨大的行业吸收和影响潜力。这些前所未有的突破为应对螺旋式上升的全球IT能源消耗提供了科学驱动的答案。作为一项深度技术，我们需要用于定量测量的专业知识和设备，以仔细建立操作机制和基准性能。该项目使用NPL在超高分辨率磁成像(QSabr NV-钻石钻机和定量磁力显微镜)和扫描热电显微镜(SThEM)中的独特设施来研究Imperial公司优化的激光写入和机器学习的纳米磁性阵列样本。该项目还可能涉及帝国理工学院的纳米制造、磁性和磁光测量。IAA的资金将使我们能够通过对Memculating标准进行基准测试来提高TRL。

We have developed an all-optical nanomagnetic writing technique using a cheap, low-power (~2 mW) laser1. Unlike other opto-magnetic approaches, it requires no magnetic field, exotic materials nor powerful lasers - making it cheap, device-friendly & industrially-scalable. In parallel, we have developed the world-first nanomagnetic array neuromorphic computing scheme, providing low power neural-network functionality2. This proof-of-concept works well but neural network weight updates rely on a slowly varying magnetic field and lack single-nanomagnet resolution. Optical writing and reading now allows us to directly write network weights (nanomagnet states). This radically enhances operation speed & solves device compatibility issues, with huge potential for industrial uptake & impact.These unprecedented breakthroughs provide science-driven answers to spiralling global IT energy consumption. As a deep technology, we need expertise and facilities for quantitative measurements to carefully establish the operating mechanisms and benchmark performance.The project is use of the unique facilities at NPL in ultra-high resolution magnetic imaging (QSabr NV-diamond rig and quantitative magnetic force microscopy) and scanning thermoelectric microscopy (sThEM) to study nanomagnetic array samples from Imperial optimised for the laser writing and machine learning. The project could also involve nanofabrication, magnetic and magneto-optic measurements at Imperial. IAA funding will allow us to raise the TRL by benchmarking against memcomputing standards.