基于物理约束与空间注意力的单分子多分类深度学习算法
Single-molecule multi-class deep learning algorithm based on physical constraints and spatial attention
摘要
[目的]针对单分子电导-位移时序数据挖掘中算法选型缺乏量化依据、深度学习模型未融合物理先验导致可解释性差、传统算法部署门槛高的问题,该文开展主流算法系统性性能对标,提出融合单分子电导先验软阈值物理约束与空间注意力的改进后卷积神经网络(CNN),并开发低门槛全流程分析工具.[方法]基于扫描隧穿裂结技术获取三类单分子电学数据,构建物理先验引导注意力聚焦电导台阶区域的CNN模型;同时开发免部署网页式程序,集成多种传统机器学习算法,实现原始数据一键式多分类分析.[结果]改进后CNN在二分类、三分类任务中准确率分别达97.7%和98.3%,较传统算法最大提升分别为27.38%与28.05%;引入不同强度高斯噪声后准确率稳定在92.6%~96.3%,抗噪性优异;注意力可视化证实模型可精准聚焦电导台阶有效区域.[结论]该文明确了不同算法的性能差异,所提模型兼具高精度、强鲁棒性与可解释性,开发的网页工具显著降低了机器学习在单分子电子学领域的应用门槛,为高通量单分子数据智能化挖掘提供了方法与工具支撑.
[Objective]In response to the lack of quantitative basis for algorithm selection in single-molecule conductance-displacement time series data mining,the poor interpretability caused by deep learning models not integrating physical priors,and the high deployment threshold of traditional algorithms,this paper conducts a systematic performance benchmarking of mainstream algorithms,proposes an improved Convolutional Neural Network(CNN)that integrates single-molecule conductance prior soft-threshold physical constraints with spatial attention,and develops a low-threshold end-to-end analysis tool.[Method]Three types of single-molecule electrical data were obtained via Scanning Tunneling Microscopy Break Junction technology to construct a classification model where physical priors guide attention to conductance plateau regions.Meanwhile,a deployment-free web-based program integrating 12 traditional algorithms was developed for one-click multi-class analysis of raw data.[Result]The improved CNN achieved accuracies of 97.7%and 98.3%in binary and ternary classification tasks,with maximum improvements of 27.38%and 28.05%over traditional algorithms,respectively.Its accuracy remained stable at 92.6%~96.3%under different intensities of Gaussian noise,demonstrating excellent noise resistance.Attention visualization confirmed that the model precisely focused on the effective regions of conductance steps.[Conclusion]This study clarifies the performance differences of various algorithms.The proposed model combines high accuracy,strong robustness and interpretability,and the developed web tool significantly lowers the application barrier of machine learning in single-molecule electronics,providing methodological and instrumental support for intelligent mining of high-throughput single-molecule data.
年,卷(期):
2026, 48(4)
分类号:
TP391
机标分类号:
TP391.41
TP18
TN911.1-34
在线出版日期:
2026年9月22日
基金项目:
2022YFB3204402:国家重点研发计划2020YFA0714703:国家重点研发计划22204135:国家自然科学基金2023JJ40619:湖南省自然科学基金项目23A0114:湖南省教育厅科学重点项目,