论文信息
题目:Physics-informed neural networks for inverse problems in nano-optics and metamaterials
作者:Yuyao Chen, Lu Lu, George Em Karniadakis, and Luca DPal Negro
期刊会议:Computational Physics
年份:19
论文地址:
代码:
内容
动机
动机:- 在强多光散射条件下,复杂多粒子几何中物理驱动的光散射微分模型的逆问题成为一个本质上不定的问题,利用传统方法不能满足预测需求- 物理信息神经网络(PINNs)是近年来发展起来的一个通用框架,用于解决偏微分方程的正问题和反问题- PINN仅使用一个训练数据集来实现所需的解决方案,从而减轻了替代方案中所需的大量训练数据集所带来的负担
问题定义:逆问题:
f
(
x
;
∂
u
^
∂
x
1
,
…
,
∂
u
^
∂
x
d
;
∂
2
u
^
∂
x
1
∂
x
1
,
…
,
∂
2
u
^
∂
x
1
∂
x
d
;
…
;
λ
)
=
0
,
x
∈
Ω
f\left(\mathbf{x} ; \frac{\partial \hat{u}}{\partial x_{1}}, \ldots, \frac{\partial \hat{u}}{\partial x_{d}} ; \frac{\partial^{2} \hat{u}}{\partial x_{1} \partial x_{1}}, \ldots, \frac{\partial^{2} \hat{u}}{\partial x_{1} \partial x_{d}} ; \ldots ; \lambda\right)=0, \quad \mathbf{x} \in \Omega
f(x;∂x1∂u^,…,∂xd∂u^;∂x1∂x1∂2u^,…,∂x1∂xd∂2u^;…;λ)=0,x∈Ω
其中
λ
\lambda
λ未知,其中loss定义为,
L
i
\mathcal{L}_{i}
Li,是初始点的
l
o
s
s
loss
loss,
L
b
\mathcal{L}_{b}
Lb是边界点的
l
o
s
s
loss
loss
L ( θ , λ ) = w f L f ( θ , λ ; T f ) + w i L i ( θ , λ ; T i ) + w b L b ( θ , λ ; T b ) \mathcal{L}(\boldsymbol{\theta}, \lambda)=w_{f} \mathcal{L}_{f}\left(\boldsymbol{\theta}, \lambda ; \mathcal{T}_{f}\right)+w_{i} \mathcal{L}_{i}\left(\boldsymbol{\theta}, \lambda ; \mathcal{T}_{i}\right)+w_{b} \mathcal{L}_{b}\left(\boldsymbol{\theta}, \lambda ; \mathcal{T}_{b}\right) L(θ,λ)=wfLf(θ,λ;Tf)+wiLi(θ,λ;Ti)+wbLb(θ,λ;Tb)其中 L f ( θ , λ ; T f ) = 1 ∣ T f ∣ ∑ x ∈ T f ∥ f ( x ; ∂ u ^ ∂ x 1 , … , ∂ u ^ ∂ x d ; ∂ 2 u ^ ∂ x 1 ∂ x 1 , … , ∂ 2 u ^ ∂ x 1 ∂ x d ; … ; λ ) ∥ 2 2 L i ( θ , λ ; T i ) = 1 ∣ T i ∣ ∑ x ∈ T i ∥ u ^ ( x ) − u ( x ) ∥ 2 2 L b ( θ , λ ; T b ) = 1 ∣ T b ∣ ∑ x ∈ T b ∥ B ( u ^ , x ) ∥ 2 2 \begin{aligned}\mathcal{L}_{f}\left(\boldsymbol{\theta}, \lambda ; \mathcal{T}_{f}\right) &=\frac{1}{\left|\mathcal{T}_{f}\right|} \sum_{\mathbf{x} \in \mathcal{T}_{f}}\left\|f\left(\mathbf{x} ; \frac{\partial \hat{u}}{\partial x_{1}}, \ldots, \frac{\partial \hat{u}}{\partial x_{d}} ; \frac{\partial^{2} \hat{u}}{\partial x_{1} \partial x_{1}}, \ldots, \frac{\partial^{2} \hat{u}}{\partial x_{1} \partial x_{d}} ; \ldots ; \lambda\right)\right\|_{2}^{2} \\\mathcal{L}_{i}\left(\boldsymbol{\theta}, \lambda ; \mathcal{T}_{i}\right) &=\frac{1}{\left|\mathcal{T}_{i}\right|} \sum_{\mathbf{x} \in \mathcal{T}_{i}}\|\hat{u}(\mathbf{x})-u(\mathbf{x})\|_{2}^{2} \\\mathcal{L}_{b}\left(\boldsymbol{\theta}, \lambda ; \mathcal{T}_{b}\right) &=\frac{1}{\left|\mathcal{T}_{b}\right|} \sum_{\mathbf{x} \in \mathcal{T}_{b}}\|\mathcal{B}(\hat{u}, \mathbf{x})\|_{2}^{2}\end{aligned} Lf(θ,λ;Tf)Li(θ,λ;Ti)Lb(θ,λ;Tb)=∣Tf∣1x∈Tf∑∥∥∥∥f(x;∂x1∂u^,…,∂xd∂u^;∂x1∂x1∂2u^,…,∂x1∂xd∂2u^;…;λ)∥∥∥∥22=∣Ti∣1x∈Ti∑∥u^(x)−u(x)∥22=∣Tb∣1x∈Tb∑∥B(u^,x)∥22
根据PINN构建如下网络:建立微分方程解的代理模型,在更加 u u u求得 l o s s loss loss,最后最小化 l o s s loss loss求得参数 θ \theta θ和 λ \lambda λ
PINNS for the homogenization of finite-size metamaterials(超材料)
(没有明确边界条件)
忽略有效介质中的辐射损失
具体应用于均质有限尺寸的超材料问题:
∇ 2 E z ( x , y ) + ε r ( x , y ) k 0 2 E z = 0 \nabla^{2} E_{z}(x, y)+\varepsilon_{r}(x, y) k_{0}^{2} E_{z}=0 ∇2Ez(x,y)+εr(x,y)k02Ez=0
其中 E z E_{z} Ez是电厂的z分量, ε r ( x , y ) \varepsilon_{r} (x, y) εr(x,y)是空间相关的相对介电常数, k = 2 π / λ 0 k=2\pi /\lambda_{0} k=2π/λ0求解 ε r ( x , y ) \varepsilon_{r} (x, y) εr(x,y)相当于参数 λ \lambda λ
实验1(阵列中散射纳米线的周期性排列):
结果:预测与真实误差7%,通过得到的电容率常数,用FEM仿真得到整个
E
z
E_{z}
Ez分布,与真实误差为2.82%
实验2(阵列中散射纳米线的周期性排列):
将实验1的
ϵ
\epsilon
ϵ改为12
结果:得到的真实与预测
E
z
E_{z}
Ez误差5%
实验3(任意非周期形态的散射阵列)
问题:沃格尔螺旋阵列(Vogel spiral array)极坐标下参数方程定义:
{
r
n
=
a
0
n
θ
n
=
n
α
\left\{\begin{array}{l}r_{n}=a_{0} \sqrt{n} \\\theta_{n}=n \alpha\end{array}\right.
{rn=a0nθn=nα
结果:误差3.8%
存在多散射介质中的辐射损失
问题:
实验:考虑存在多散射介质中的辐射损失应用于前面1,2,3实验中,得到 Im { ε r ( x , y ) } \operatorname{Im}\left\{\varepsilon_{r}(x, y)\right\} Im{εr(x,y)}为 1 0 − 4 10^{-4} 10−4,0.6和0.3, Re { ε r ( x , y ) } \operatorname{Re}\left\{\varepsilon_{r}(x, y)\right\} Re{εr(x,y)}误差非常都接近,3%
创新:论文中提出的一般的PINNs框架适用于研究任意非周期有效介质和超材料中形态和辐射耦合的影响,超越了有效介质理论的限制。
PINN for inverse MIE scattering (有明确的边界条件)
创新:PINNs框架可以用于从近场成像数据中检索复杂纳米结构的光学特性
实验1:
问题:
{
∇
2
E
z
(
k
)
+
ε
r
k
k
0
2
E
z
(
k
)
=
0
in
Ω
k
,
(
k
=
1
,
2
)
E
z
(
1
)
∣
r
=
a
=
E
z
(
2
)
∣
r
=
a
1
μ
r
1
∂
E
z
(
1
)
∂
r
∣
r
=
a
=
1
μ
r
2
∂
E
z
(
2
)
∂
r
∣
r
=
a
\left\{\begin{array}{l}\nabla^{2} E_{z}^{(k)}+\varepsilon_{r k} k_{0}^{2} E_{z}^{(k)}=0 \quad \text { in } \Omega_{k},(k=1,2) \\\left.E_{z}^{(1)}\right|_{r=a}=\left.E_{z}^{(2)}\right|_{r=a} \\\left.\frac{1}{\mu_{r 1}} \frac{\partial E_{z}^{(1)}}{\partial r}\right|_{r=a}=\left.\frac{1}{\mu_{r 2}} \frac{\partial E_{z}^{(2)}}{\partial r}\right|_{r=a}\end{array}\right.
⎩⎪⎪⎨⎪⎪⎧∇2Ez(k)+εrkk02Ez(k)=0 in Ωk,(k=1,2)Ez(1)∣∣∣r=a=Ez(2)∣∣∣r=aμr11∂r∂Ez(1)∣∣∣r=a=μr21∂r∂Ez(2)∣∣∣r=a
其中
E
z
(
1
)
,
E
z
(
2
)
E_{z}^{(1)}, E_{z}^{(2)}
Ez(1),Ez(2)表示电场z分量的分别nanocylinder的内部和外部实部,实验中设置
ε
r
2
=
1
\varepsilon_{r 2}=1
εr2=1而
ε
r
1
\varepsilon_{r 1}
εr1为需要训练的参数
结果:误差0.51%
实验2:多参数问题:
{
∇
2
E
z
(
k
)
+
ε
k
k
0
2
E
z
(
k
)
=
0
in
Ω
k
,
(
k
=
1
,
2
,
3
)
E
z
(
1
)
∣
r
=
a
=
E
z
(
2
)
∣
r
=
a
,
E
z
(
2
)
∣
r
=
b
=
E
z
(
3
)
∣
r
=
b
1
μ
i
∂
E
z
(
1
)
∂
r
∣
r
=
a
=
1
μ
c
∂
E
z
(
2
)
∂
r
∣
r
=
a
,
1
μ
c
∂
E
z
(
2
)
∂
r
∣
r
=
b
=
1
μ
0
∂
E
z
(
3
)
∂
r
∣
r
=
b
\left\{\begin{array}{l}\nabla^{2} E_{z}^{(k)}+\varepsilon_{k} k_{0}^{2} E_{z}^{(k)}=0 \quad \text { in } \Omega_{k},(k=1,2,3) \\\left.E_{z}^{(1)}\right|_{r=a}=\left.E_{z}^{(2)}\right|_{r=a},\left.\quad E_{z}^{(2)}\right|_{r=b}=\left.E_{z}^{(3)}\right|_{r=b} \\\left.\frac{1}{\mu_{i}} \frac{\partial E_{z}^{(1)}}{\partial r}\right|_{r=a}=\left.\frac{1}{\mu_{c}} \frac{\partial E_{z}^{(2)}}{\partial r}\right|_{r=a},\left.\quad \frac{1}{\mu_{c}} \frac{\partial E_{z}^{(2)}}{\partial r}\right|_{r=b}=\left.\frac{1}{\mu_{0}} \frac{\partial E_{z}^{(3)}}{\partial r}\right|_{r=b}\end{array}\right.
⎩⎪⎪⎨⎪⎪⎧∇2Ez(k)+εkk02Ez(k)=0 in Ωk,(k=1,2,3)Ez(1)∣∣∣r=a=Ez(2)∣∣∣r=a,Ez(2)∣∣∣r=b=Ez(3)∣∣∣r=bμi1∂r∂Ez(1)∣∣∣r=a=μc1∂r∂Ez(2)∣∣∣r=a,μc1∂r∂Ez(2)∣∣∣r=b=μ01∂r∂Ez(3)∣∣∣r=b
结论
- 成功地引入并验证了PINNs框架用于有效的散射阵列介质重建
不懂
- 里面一些物理背景
- 逆问题求解 λ \lambda λ,是与x,y有关的参数