作者介绍:
Division of Applied Mathematics, Brown University 布朗大学应用数学系
LESC,ICMSEC
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image.png)
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 1.png)
第一作者 Lu Lu PINN的研究,以及红细胞的相关研究
ORCID(全球科研人员的学术身份证):
https://orcid.org/0000-0002-5476-5768
George Em Karniadakis (布朗大学) PDE、PINN、deepOnet的相关研究,以及物理和生物学系统**physical and biological systems (especially the brain)(from https://www.cfm.brown.edu/faculty/gk/),**以下是他讲解deepOnet的视频
https://www.bilibili.com/video/BV15341177x2/?vd_source=83444af567ef1cd68516b404a00da866
下面的链接是讲解此论文的,但是比较难懂
前提条件:
范数是一种将向量空间中的向量映射到非负实数的函数,Banach空间是一个完备的赋范空间,with norm应该是代表“赋范”过程
compact set紧集 在这个集合中,无论怎么选择一个无限序列,总能找到一个部分序列,它最终会收敛于集合内的某个点。
输入函数u(x) 输出函数s(x) 在实践中离散为信号
算子:输入到输出函数空间的一个映射
f:函数或泛函
G:算子(operator)
activation function 激活函数
sensor point 训练的数据点
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 2.png)
主要内容:
目的:学习得到算子G
(与算子相关的)无限逼近定理Universal Approximation Theorem for Operator
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 3.png)
![5a1e84ec2a9eb90060b978d42ab32b1.jpg](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/5a1e84ec2a9eb90060b978d42ab32b1.jpg)
最主要的图:
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 4.png)
图1:输入分为两部分, 分别是u(x1)……u(xm) y
放入一个网络,输出G(u)(y)
这个网络可以选择传统网络FNN,把 [ u ( x 1 ) … … u ( x m ) , y ] T [u(x_1)……u(x_m),y]^\mathsf{T} [u(x1)……u(xm),y]T直接放进去
![4328969b07c6e8c8541fc1ff2786218.jpg](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/4328969b07c6e8c8541fc1ff2786218.jpg)
图3 p个分支网络并行堆叠,“堆叠DeepONet”
考虑到y和u(xi)维数不同,用两个网络分别处理,y通过主干网络输出p个t,u(xi)经过p个分支输出b,根据Theorem 1得到近似:
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 5.png)
虽然Theorem 1未提到偏置,但增加偏置可以提高泛化性能,如下式:
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 6.png)
图4“非堆叠DeepONet”
p>10,而实践中大量分支网络太过占用计算和内存,于是将p个分支网络合并成一个,单个网络输出p个b
采集数据:从两个函数空间高斯随机场(GRF)和正交(切比雪夫)多项式
高斯随机场(GRF)
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 7.png)
正交(切比雪夫)多项式,在[-M,M]随机采样ai得到u
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 8.png)
用龙格-库塔(4,5)法求解ODE系统,用二阶有限差分法求解PDEs得到参考解。则一个数据点是一个三元组(u, y, G(u)(y))
结果:得到了数个u,每个u对应数个y和G(u)(y)
一个采样的实例:动力学系统
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 9.png)
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 10.png)
G(u)就是s,都是output function,对下式求导/令x=a就得到上式
采样x,定义u
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 11.png)
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 12.png)
定义Um,Wm,W等紧集,得到这样的近似理论
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 13.png)
结果
第一个例子:一维动力学系统
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 14.png)
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 15.png)
depth from 2 to 4, width from 10 to 2560, and learning rate from 0.0001 to 0.01
-
深度(Depth):指神经网络的层数。从2到4的深度意味着网络将有2至4层的隐藏层。
-
宽度(Width):指每一层中神经元的数量。从10到2560的宽度范围意味着在每一层中,神经元的数量会在10到2560之间变化。
-
学习率(Learning Rate):是训练神经网络时用来更新权重的参数,决定了每次参数更新的步长。(可更新的参数一般是权重(weights)和偏置(biases))
-
迭代(Iterations): 使用一批数据使用一个批次(batch)的数据进行前向传播、计算损失函数、反向传播,并更新参数。一个周期(Epoch)通常包含多次迭代。
FNN与DeepOnet比较:
FNNs的结果D2W10就是深度2层宽度10个神经元,从上到下,学习率分别为0.01 0.001 0.0001,阴影是10组改变条件(训练数据、初始条件)的数据集(10 runs with random training dataset and network initialization)得到的一些SD
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 16.png)
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 17.png)
(A)带有偏差的未堆叠DeepONet的训练轨迹 (B)各种对比,使用的是10个改变条件(训练数据、初始条件)数据集
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 18.png)
是否堆叠的比较:
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 19.png)
A 随着Training,是否堆叠的训练和测试误差都变小,
B 10个改变条件(训练数据、初始条件)的数据集 的结果
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 20.png)
C MSE 非堆叠测试机MSE小,泛化性好
得出结论,使用非堆叠的方法
第二个例子:
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 21.png)
This problem is characterized by three factors:
(1) k,
(2) maximum prediction time T,
(3) input function space.
The accuracy of learned networks is determined by four factors:
(1) the number of sensor points m;
(2) training dataset size;
(3) network architecture,
(4) optimizer.
随着sensors增加,误差趋于稳定。BCD是T l k对需要的sensors的影响,sensors与T数目正比,l反比,k基本不影响。
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 22.png)
示例三:扩散-反应系统
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 23.png)
D是扩散系数,学习从u到s的算子G
数据集的采样:
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 24.png)
对于某一个u,采样x和u(x) ——采样输出函数s——获得p对数据点,每个点都是一个三元组
会采样多种u,最终数据集数据量是p*u的个数
固定u或P,MSE随另一个因素的变化: 阴影区域表示单标准差
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 25.png)
ABDE 在收敛之前,P与u的MSE呈一种多项式或者指数变化的关系;CE 是多项式或者指数拟合的参数随P和u变化的拟合。
![image.png](Lu等的DeepONet:+Learning+nonlinear+operators+for+identifying+differential
equations+based+on+the+universal+approximation+theorem+of
operators+23899b1f-03b4-4f50-9749-30087e442897/image 26.png)
拓展资料:
讲文章的视频:
详细讲解了Notations的内容,遇到很多新概念。
中文注解,有助于理解: