基于物理信息机器学习的酶促反应系统参数估计  

Parameter estimation of enzymatic reaction systems based on physics-informed machine learning

作　　者：刘承杰 俞辉 陈宇 戴厚德 LIU Chengjie;YU Hui;CHEN Yu;DAI Houde(School of Advanced Manufacturing,Fuzhou University,Quanzhou 362000,Fujian Province,China;Fujian Institute of Research on the Structure of Matter,Chinese Academy of Sciences,Fuzhou 350108,Fujian Province,China;College of Mechanical and Electrical Engineering,Fujian Agriculture and Forestry University,Fuzhou 350100,Fujian Province,China;Fujian Key Laboratory of Special Intelligent Equipment Safety Measurement and Control,Fujian Special Equipment Inspection and Research Institute,Fuzhou 350008,Fujian Province,China)

机构地区：[1]福州大学先进制造学院,福建泉州362000 [2]中国科学院福建物质结构研究所,福建福州350108 [3]福建农林大学机电工程学院,福建福州350100 [4]福建省特种设备检验研究院福建省特种智能装备安全与测控重点实验室,福建福州350008

出　　处：《化学工程》2024年第7期77-81,94,共6页Chemical Engineering(China)

基　　金：福建省特种智能装备安全与测控重点实验室(FJIES2023KF02);泉州市科技计划项目(2022C004L)。

摘　　要：为揭示物理信息神经网络在生化领域中的潜力,研究一种基于现代物理信息机器学习工具的新参数估计方法,并通过酶促反应过程模型的案例研究进行了演示,比较软、硬边界约束设置对计算结果的影响。实验分析表明,利用软、硬2种不同约束的物理信息神经网络均能获得精确的模型参数估计值,并在所有的可观测变量上的拟合优度R^(2)在0.98以上,所得到的系统模型能够较好地反映系统的动态过程。所提出的方法融合了模型驱动与数据驱动方法的优势,并且能够在基于采样40次的含噪声小型数据集上获得稳健的训练结果,显著降低对数据量的要求。To reveal the potential of physics-informed neural networks in biochemistry,a new parameter estimation method based on modern physics-informed machine learning tools was investigated and its function was demonstrated through a case study of enzymatic synthesis process and the effects of soft and hard boundary constraint settings were compared on the computational results.The experimental results show that both physics-informed neural networks with soft and hard constraints can accurately estimate model parameters,with goodness of fit R^(2)above 0.98 on all observable variables.The resulting system model can better reflect the dynamic process of the system.The proposed method combines the advantages of model-driven and data-driven approaches and achieves robust training results on a small dataset based on 40 noisy samples,significantly reducing the required data.

关 键 词：物理信息嵌入 酶促反应 神经网络 参数估计 硬约束 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程] TQ033[自动化与计算机技术—控制科学与工程]