A Machine Learning Approach to Support Neuromorphic Device Design and Microfabrication
A Machine Learning Approach to Support Neuromorphic Device Design and Microfabrication
复制标题
DOI：
10.1109/icmla58977.2023.00246
复制
发表时间：
2023-12
期刊：
影响因子：
--
通讯作者：
A. Vicenciodelmoral;Md. Mehedi Hasan Tanim;Feng Zhao;Xinghui Zhao
中科院分区：
文献类型：
--
作者：
A. Vicenciodelmoral;Md. Mehedi Hasan Tanim;Feng Zhao;Xinghui Zhao
神经形态芯片为可持续计算提供了潜在的解决方案，因为它们试图模仿人脑中的neu-ronal架构，并在减少数量级的能源消耗方面具有巨大的潜力并改善计算性能。但是，神经芯片的制造过程是昂贵的，目前是基于反复试验的，这增加了设计过程的复杂性。在本文中，我们通过设计和开发机器学习指导的微加工过程来解决这些挑战，该过程是电阻随机访问记忆（RRAM），这是神经形态芯片中的关键设备。具体而言，我们的研究做出了以下贡献：1）我们使用生物有机材料成功地制造了一种新的RRAM，从而为支持神经形态计算提供了更绿色，更可持续的解决方案； 2）我们对微加工过程条件及其对RRAM设备的影响进行了全面研究，从而为该领域提供了新知识。 3）我们开发了一种合成数据辅助方法，以预测生物有机RRAM（Bio-RRAM）设备的关键性能指标，而无需大量的实验培训数据； 4）我们开发了一种更高级的方法，该方法利用学习任务转换来对细粒度的性能指标进行预测，而对实验或合成数据没有附加要求。我们使用我们制造的八种基于蜂蜜的生物rrAM设备评估了这些方法，结果表明，这两种方法在预测设备性能方面都是有效的。我们预计机器学习引导的微加工将为绿色神经形态计算的下一代RRAM设备的下一代设计铺平道路。
Neuromorphic chips provide a potential solution for sustainable computing, as they attempt to mimic the neu-ronal architectures in human brain and show great potentials in reducing energy consumption in the order of magnitude and also improve the computational performance. However, the fabrication process for neuromophic chips is costly and currently based on trial-and-error, which adds complexity to the design process. In this paper, we address these challenges by designing and developing machine learning guided microfabrication process for Resistive Random Access Memory (RRAM), which is a key device in neuromorphic chips. Specifically, our research makes the following contributions: 1) we successfully fabricated a new RRAM using bio-organic materials, leading to a greener and more sustainable solution for supporting neuromorphic computing; 2) we carried out a comprehensive study on the microfabrication process conditions and their effects on the RRAM devices, producing new knowledge to the field; 3) we developed a synthetic data assisted approach to predict key performance metrics of the bio-organic RRAM (bio-RRAM) devices, without requiring substantial amount of experimental training data; and 4) we developed a more advanced approach which leverages learning task conversion to make predictions on fine-grained performance metrics with no added requirements for experimental or synthetic data. We evaluated these approaches using eight honey-based bio-RRAM devices we fabricated, and the results show that both approaches are effective in terms of predicting the devices performance. We expect that the machine learning guided microfabrication will pave the way to more efficient and effective design of the next-generation of RRAM devices for green neuromorphic computing.