2025.2.9机器学习笔记：PINN文献阅读

文献阅读

题目信息

• 题目： GPT-PINN:Generative Pre-Trained Physics-Informed Neural Networks toward non-intrusive Meta-learning of parametric PDEs
• 期刊： Finite Elements in Analysis and Design
• 作者： Yanlai Chen, Shawn Koohy
• 发表时间： 2024
• 文章链接： https://arxiv.org/pdf/2303.14878

摘要

PINN是一种结合了深度神经网络和现代高性能计算技术的方法，其可以用来求解复杂的非线性偏微分方程。虽然PINN能够处理复杂的方程，但在需要多次计算或实时模拟的情况下它的训练过程很慢。此外，PINN的结构通常比较复杂，参数太多，导致计算效率不高。传统的数值方法虽然也能解决这些问题，但计算量非常大。为了解决PINN的这些缺点，这篇论文提出了一种新的方法，叫做生成式预训练PINN（GPT-PINN）。GPT-PINN的核心思想是简化网络结构，减少计算量。此外，GPT-PINN的外部结构非常简单，只有一个隐藏层，而且隐藏层中的神经元数量也大大减少。其每个神经元的激活函数是通过预先训练好的完整PINN来确定的，这些预训练是在一些特定选择的系统配置下完成的。GPT-PINN还引入了一个“元网络”，这个元网络能够自适应地学习系统的参数变化，并且每次只增加一个神经元。通过这种方式，GPT-PINN只需要训练少量的网络，就能在整个参数范围内快速生成准确的解。这种方法不仅提高了计算效率，还能更好地理解系统在不同参数下的行为。

Abstract

Physics-Informed Neural Networks (PINNs) combine deep learning with high-performance computing to solve nonlinear PDEs but face challenges like slow training and excessive parameterization. This paper introduces Generative Pre-trained PINN (GPT-PINN), which simplifies the network to a single hidden layer with fewer neurons. Each neuron’s activation function is pre-trained under specific configurations, and a meta-network adaptively learns parameter dependencies, adding neurons incrementally. GPT-PINN efficiently generates accurate solutions across the parameter domain with minimal training, improving computational efficiency and system understanding. This approach is particularly useful for multi-query applications like uncertainty quantification and optimal control.

创新点

1.网络架构设计，将预训练的完整网络作为神经元的激活函数；
2.采用元网络的训练损失作为误差指标，灵感源于传统数值求解器基于残差的误差估计。

网络架构

这幅图展示了GPT-PINN的架构，该架构受到降阶基方法（RBM） 的启发，主要用于求解参数化偏微分方程
RBM
RBM的关键在于假设所有参数值对应的解 $u(\cdot ;\mu )$ 可以在一个较小的子空间 $X_N$​ 中很好地近似。这个子空间的大小由Kolmogorov N-width决定，它衡量了解集在最佳N维子空间中的最大误差。
算法流程如下：
1.从一组参数值 ${\mu }_1$​ 开始，求解全阶模型（FOM），并将第一个解 $u(\cdot ;\mu 1)$ 作为初始降阶基 $X_1$​。
2.使用贪心算法逐步增加降阶基 $X_n$​。对于每个参数值 $\mu$，求解降阶模型（ROM），计算误差估计${\Delta }_{n-1}(\mu )$，选择使误差最大的参数值 ${\mu }_n$​，求解FOM，更新降阶基 $X_n$​。
3.当满足某个条件，比如误差低于某个阈值或达到最大基函数数量时，停止算法。
Kolmogorov N-width
Kolmogorov N-width $d_N$​ 衡量了解集在最佳N维子空间中的最差近似误差。
公式如下：
$$d_N[u(\cdot ;D)]=\underset{X_N\subset X_h,\dim X_N=N}{\inf }\underset{\mu \in D}{\sup }\underset{v\in X_N}{\inf }\Vert u(\cdot ,\mu )-v{\Vert }_X$$

第一个$inf$表示在所有可能的N维子空间 $X_N$​ 中寻找最佳的一个。
$sup$表示考虑所有参数 $\mu$，找到最差情况下的误差。
最后一个$inf$表示在选定的子空间 $X_N$​ 中找到每个 $u(\cdot ,\mu )$ 的最佳近似 v。

降阶基
降阶基是一组能够代表高维解空间的基函数。通过这些基函数的线性组合，我们可以近似解 $u(\cdot ;\mu )$。

以下是对该架构的详细说明：

1. 超降阶神经网络：图中的神经网络被称为超降阶网络（hyper-reduced），其通过大幅简化传统神经网络结构，从而降低计算量并加快学习速度。该网络用$NN(2,n,1)$表示，其中2表示输入层节点数（空间位置x和时间t），n为隐藏层节点数，1为输出层节点数（温度u）。

    其中“超降阶”则是模型降阶中的一种技术，可能进一步简化计算步骤，比如减少参数空间或优化计算过程。在GPT-PINN的上下文中，元网络（meta-network）被设计为超降阶的，意味着它的结构比传统的PINN更简单例如只有一个隐藏层，神经元数量减少，并且每个神经元使用预训练的完整PINN作为激活函数。“超降阶”中的“超”不是指超越，而是指进一步的降阶或优化。比如，传统的降阶方法可能减少变量数目，而超降阶可能在降阶的基础上进一步优化计算流程或网络结构，使其更加高效。
   

2. 隐藏层激活函数：隐藏层的每个节点都配备了一个特定的激活函数，这些函数实际上是针对特定参数值${\mu }^n$预训练的PINN。即每个隐藏层节点并非简单的权重和偏置组合，而是一个完整的预训练网络，能够根据物理定律和已知的PDEs对输入做出响应。

3. 贪婪算法选择的参数：参数值{${\mu }^1,{\mu }^2,...,{\mu }^n$}是通过专门针对PINNs设计的贪婪算法选取的，以确保所选参数能够有效捕捉解的整体特征。

    贪婪算法是一种“每一步都选当前最优解”的策略，不求全局最优，但求快速逼近。
   

贪婪算法在GPT-PINN的运作过程如下图所示：
1.从一个随机选择的 μ^1开始，设置一个空的 GPT-PINN，n=1。
2.在 μⁿ处训练 GPT-PINN 并求解完整的 PINN。
3.将新的神经元添加到 GPT-PINN 中。
4.检查所有训练参数的最大终端损失是否低于容差${\epsilon }_{tol}$。
5.若满足条件，则训练过程结束；如果未收敛，选择下一个最差的 μⁿ⁺¹，将 n 增加 1，然后重复。
其核心思想就是通过逐步添加神经元来训练GPT-PINN，每次选择使当前模型损失最大的参数进行训练。目标是确保所有训练参数的损失都低于某个容差，从而确保模型的泛化能力。

4. 输出：输出层将所有隐藏层节点的输出进行整合，从而生成给定参数值下空间位置x和时间t处的温度u的近似解。
   GPT-PINN的解是多个预训练PINN的线性组合，其公式表达如下：
   $$u_{\mathrm{GPT}}(\mathbf{x},t;\mathbf{\mu })=\sum _{i=1}^n{c}_i(\mathbf{\mu })\cdot {\Psi }_{\mathrm{NN}}^{{\theta }^i}(\mathbf{x},t)$$
   参数的含义如下：
   x为空间坐标；t为时间变量；μ为参数向量（如材料属性、边界条件等）；
   ${\sum }_{i=1}^n$表示对n个预训练的PINN模型进行线性组合；
   $c_i(\mathbf{\mu })$表示第i个预训练PINN的权重系数，依赖参数 μ；
   ${\Psi }_{\mathrm{NN}}^{{\theta }^i}(\mathbf{x},t)$表示第 i个预训练的PINN，参数 θⁱ 在特定参数 μⁱ下训练完成。

实验

这篇论文将生成式预训练物理信息神经网络（GPT - PINN）应用于三个方程族（Klein - Gordon方程、Burgers方程和Allen - Cahn方程）进行数值实验。
对于这些方程的具体表达，可以查看我之前的文章：第二十六周机器学习笔记：PINN求正反解求PDE文献阅读

实验结果如下：

1. 对Klein - Gordon方程的实验结果
训练过程：
全连接网络结构的全PINN采用特定激活函数、学习率、优化器、轮数等进行训练。参数训练集有1000个值，贪心算法生成最多15个神经元的GPT - PINN，其在与全PINN相同的配点集上用不同学习率和轮数训练。 GPT - PINN生成的15个神经元对应的参数值及离线训练损失显示，采样参数值靠近域边界，训练损失呈指数下降。与均匀采样对比，自适应“学习神经元”性能更好，且达到一定精度只需训练全PINN 15次。

测试过程：
在200个随机选择的参数值上测试GPT - PINN，不同大小的GPT - PINN最大误差呈指数趋势。对比全PINN和GPT - PINN的累积运行时间，GPT - PINN边际成本低，增长缓慢，约为全PINN的0.0022倍。全PINN和GPT - PINN的训练损失随轮数变化，GPT - PINN损失下降更平滑，还给出了GPT - PINN解的逐点误差。

2. 对Burgers方程的实验结果
训练过程：
全PINN为特定结构的全连接网络参数训练集在粘度域内均匀分布，贪心算法生成最多9个神经元的GPT - PINN，其在排除部分配点（靠近大梯度处）的相同配点集上用不同学习率和轮数训练。GPT - PINN生成的9个神经元对应的参数值及离线训练损失显示出与Klein - Gordon方程类似的结果，自适应“学习神经元”比非自适应“均匀神经元”性能好很多。

测试过程：
在25个参数值上测试GPT - PINN，不同大小的GPT - PINN最大误差呈指数趋势。对比全PINN和GPT - PINN的累积运行时间，GPT - PINN增长缓慢（相对速度为0.009），少量查询（12次）时投资GPT - PINN就很值得。全PINN和GPT - PINN的训练损失随轮数变化，GPT - PINN损失下降更平滑，验证了初始化策略的有效性。

3. 对Allen - Cahn方程的实验结果
训练过程：
采用自适应性的SA - PINN作为全PINN，设定网络结构、激活函数、学习率、优化器、轮数等进行训练。参数训练集为均匀分布的121个值，贪心算法生成最多9个神经元的GPT - PINN，其在与SA - PINN相同的配点集上用不同学习率和轮数训练。
GPT - PINN生成的9个神经元对应的参数值及离线训练损失显示出与前两个方程类似的结果。

测试过程：
在25个参数值上测试GPT - PINN，不同大小的GPT - PINN最大误差及累积运行时间表明，GPT - PINN增长缓慢（相对速度为0.0006），少量查询（9 - 10次）时训练GPT - PINN就很值得。

SA - PINN和GPT - PINN的训练损失随轮数变化，GPT - PINN损失下降更平滑。

代码链接：https://github.com/idrl-lab/PINNpapers

代码如下：

# Import and GPU Support
import matplotlib.pyplot as plt
import numpy as np
import torch
import os
import tensorflow as tfphysical_devices = tf.config.list_physical_devices('GPU') 
tf.config.experimental.set_memory_growth(physical_devices[0], True)import time
import scipy.io
from tensorflow import keras
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Dense, Input
from tensorflow.keras import layers, activations
from eager_lbfgs import lbfgs, Struct
from pyDOE import lhs
np.random.seed(1234) # lhs# GPT-PINN
from AC_Plotting import AC_plot, loss_plot from AC_GPT_activation import P
from AC_GPT_precomp import autograd_calculations, Pt_lPxx_eP
from AC_GPT_PINN import GPT
from AC_GPT_train import gpt_train# SA-PINN is implemented explicitly in the codetorch.set_default_dtype(torch.float)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Current Device (PyTorch): {device}")
if torch.cuda.is_available():print(f"Current Device Name (PyTorch): {torch.cuda.get_device_name()}")# Domain and Data
data_path = r".\data"
xt_test = torch.from_numpy(np.loadtxt(fr"{data_path}\xt_test.txt",dtype=np.float32)).to(device)# Initial Condition
IC_x = torch.from_numpy(np.loadtxt(fr"{data_path}\initial\x0.txt",dtype=np.float32))
IC_t = torch.zeros(IC_x.shape[0])[:,None]
val, idx = torch.sort(IC_x)
IC_u = torch.from_numpy(np.loadtxt(fr"{data_path}\initial\u0.txt",dtype=np.float32))
IC_u = IC_u[idx][:,None].to(device)
IC_xt = torch.hstack((IC_x[idx][:,None],IC_t)).to(device)# Boundary Condition
BC_x_ub = torch.from_numpy(np.loadtxt(fr"{data_path}\boundary\x_ub.txt",dtype=np.float32))[:,None]
BC_t_ub = torch.from_numpy(np.loadtxt(fr"{data_path}\boundary\t_ub.txt",dtype=np.float32))[:,None]
BC_xt_ub = torch.hstack((BC_x_ub,BC_t_ub))BC_x_lb = torch.from_numpy(np.loadtxt(fr"{data_path}\boundary\x_lb.txt",dtype=np.float32))[:,None]
BC_t_lb = torch.from_numpy(np.loadtxt(fr"{data_path}\boundary\t_lb.txt",dtype=np.float32))[:,None]
BC_xt_lb = torch.hstack((BC_x_lb,BC_t_lb))BC_xt = torch.vstack((BC_xt_ub,BC_xt_lb)).to(device) 
BC_u = torch.full((200,1), -1.0).to(device)  # Residual 
x_resid = torch.from_numpy(np.loadtxt(fr"{data_path}\f\xf_train.txt",dtype=np.float32))[:,None]
t_resid = torch.from_numpy(np.loadtxt(fr"{data_path}\f\tf_train.txt",dtype=np.float32))[:,None]xt_resid = torch.hstack((x_resid,t_resid)).to(device) 
f_hat = torch.full((20000,1), 0.0).to(device)   # Training Parameter Set
ac_training = np.loadtxt("ac_param_training.txt")train_final_gpt   = True
number_of_neurons = 9
loss_list         = np.ones(number_of_neurons)
print(f"Expected Final GPT-PINN Depth: {[2,number_of_neurons,1]}\n")###############################################################################
#################################### GPT Setup ####################################
###############################################################################
P_resid_values    = torch.ones((xt_resid.shape[0],  number_of_neurons)).to(device)
P_IC_values       = torch.ones((   IC_xt.shape[0],  number_of_neurons)).to(device)
P_BC_values       = torch.ones((   BC_xt.shape[0],  number_of_neurons)).to(device)P_t_term  = torch.ones((xt_resid.shape[0], number_of_neurons)).to(device)
P_xx_term = torch.ones((xt_resid.shape[0], number_of_neurons)).to(device)ac_neurons    = [1 for i in range(number_of_neurons)]
ac_neurons[0] = [0.00055, 3.0]P_list = np.ones(number_of_neurons, dtype=object)lr_gpt          = 0.0025
epochs_gpt      = 2000
epochs_gpt_test = 5000
test_cases      = 25# Save Data/Plot Options
save_data         = False
plot_pinn_loss    = True
plot_pinn_sol     = True
plot_largest_loss = Truepinn_train_times = np.ones(number_of_neurons)
gpt_train_times  = np.ones(number_of_neurons)total_train_time_1 = time.perf_counter()
###############################################################################
#################################### SA Setup #################################
###############################################################################
# SA-PINN from https://github.com/levimcclenny/SA-PINNslr_adam_sa  = 0.005
lr_lbfgs_sa = 0.8epochs_adam_sa  = 10000
epochs_lbfgs_sa = 10000layers_sa = [2, 128, 128, 128, 128, 1]lb = np.array([-1.0])
ub = np.array([1.0])N0 = 512
N_b = 100
N_f = 20000data = scipy.io.loadmat('AC.mat')
t = data['tt'].flatten()[:,None]
x = data['x'].flatten()[:,None]
Exact = data['uu']
Exact_u = np.real(Exact)#grab training points from domain
idx_x = np.random.choice(x.shape[0], N0, replace=False)
x0 = x[idx_x,:]
u0 = tf.cast(Exact_u[idx_x,0:1], dtype = tf.float32)idx_t = np.random.choice(t.shape[0], N_b, replace=False)
tb = t[idx_t,:]
# Grab collocation points using latin hypercube sampling
X_f = lb + (ub-lb)*lhs(2, N_f)
x_f = tf.convert_to_tensor(X_f[:,0:1], dtype=tf.float32)
t_f = tf.convert_to_tensor(np.abs(X_f[:,1:2]), dtype=tf.float32)# IC/BC data 
X0 = np.concatenate((x0, 0*x0), 1) # (x0, 0)
X_lb = np.concatenate((0*tb + lb[0], tb), 1) # (lb[0], tb)
X_ub = np.concatenate((0*tb + ub[0], tb), 1) # (ub[0], tb)
x0 = tf.cast(X0[:,0:1], dtype = tf.float32)
t0 = tf.cast(X0[:,1:2], dtype = tf.float32)
x_lb = tf.convert_to_tensor(X_lb[:,0:1], dtype=tf.float32)
t_lb = tf.convert_to_tensor(X_lb[:,1:2], dtype=tf.float32)
x_ub = tf.convert_to_tensor(X_ub[:,0:1], dtype=tf.float32)
t_ub = tf.convert_to_tensor(X_ub[:,1:2], dtype=tf.float32)#L-BFGS weight getting and setting from https://github.com/pierremtb/PINNs-TF2.0
def set_weights(model, w, sizes_w, sizes_b):with tf.device('/GPU:0'):for i, layer in enumerate(model.layers[0:]):start_weights = sum(sizes_w[:i]) + sum(sizes_b[:i])end_weights = sum(sizes_w[:i+1]) + sum(sizes_b[:i])weights = w[start_weights:end_weights]w_div = int(sizes_w[i] / sizes_b[i])weights = tf.reshape(weights, [w_div, sizes_b[i]])biases = w[end_weights:end_weights + sizes_b[i]]weights_biases = [weights, biases]layer.set_weights(weights_biases)def get_weights(model):with tf.device('/GPU:0'):w = []for layer in model.layers[0:]:weights_biases = layer.get_weights()weights = weights_biases[0].flatten()biases = weights_biases[1]w.extend(weights)w.extend(biases)w = tf.convert_to_tensor(w)return w#define the neural network model
def neural_net(layer_sizes):with tf.device('/GPU:0'):model = Sequential()model.add(layers.InputLayer(input_shape=(layer_sizes[0],)))for width in layer_sizes[1:-1]:model.add(layers.Dense(width, activation=tf.nn.tanh,kernel_initializer="glorot_normal"))model.add(layers.Dense(layer_sizes[-1], activation=None,kernel_initializer="glorot_normal"))return model#define the loss
def loss(x_f_batch, t_f_batch,x0, t0, u0, x_lb,t_lb, x_ub, t_ub,col_weights, u_weights, lmbda, eps):with tf.device('/GPU:0'):f_u_pred = f_model(x_f_batch, t_f_batch, lmbda, eps)u0_pred = u_model(tf.concat([x0, t0], 1))u_lb_pred, u_x_lb_pred, = u_x_model(u_model, x_lb, t_lb)u_ub_pred, u_x_ub_pred, = u_x_model(u_model, x_ub, t_ub)mse_0_u = tf.reduce_mean(tf.square(u_weights*(u0 - u0_pred)))mse_b_u = tf.reduce_mean(tf.square(tf.math.subtract(u_lb_pred, u_ub_pred))) + \tf.reduce_mean(tf.square(tf.math.subtract(u_x_lb_pred, u_x_ub_pred)))mse_f_u = tf.reduce_mean(tf.square(col_weights * f_u_pred[0]))return  mse_0_u + mse_b_u + mse_f_u , tf.reduce_mean(tf.square((u0 - u0_pred))), mse_b_u, tf.reduce_mean(tf.square(f_u_pred))@tf.function
def f_model(x,t,lmbda,eps):with tf.device('/GPU:0'):u = u_model(tf.concat([x, t],1))u_x = tf.gradients(u, x)u_xx = tf.gradients(u_x, x)u_t = tf.gradients(u,t)c1 = tf.constant(lmbda, dtype = tf.float32)c2 = tf.constant(eps, dtype = tf.float32)f_u = u_t - c1*u_xx + c2*u*u*u - c2*ureturn [e-mail]
def u_x_model(u_model, x, t):with tf.device('/GPU:0'):u = u_model(tf.concat([x, t],1))u_x = tf.gradients(u, x)return u, [e-mail]
def grad(model, x_f_batch, t_f_batch, x0_batch, t0_batch, u0_batch, x_lb, t_lb, x_ub, t_ub, col_weights, u_weights, lmbda, eps):with tf.device('/GPU:0'):with tf.GradientTape(persistent=True) as tape:loss_value, mse_0, mse_b, mse_f = loss(x_f_batch, t_f_batch, x0_batch, t0_batch, u0_batch, x_lb, t_lb, x_ub, t_ub, col_weights, u_weights, lmbda, eps)grads = tape.gradient(loss_value, u_model.trainable_variables)grads_col = tape.gradient(loss_value, col_weights)grads_u = tape.gradient(loss_value, u_weights)gradients_u = tape.gradient(mse_0, u_model.trainable_variables)gradients_f = tape.gradient(mse_f, u_model.trainable_variables)return loss_value, mse_0, mse_b, mse_f, grads, grads_col, grads_u, gradients_u, gradients_fdef fit(x_f, t_f, x0, t0, u0, x_lb, t_lb, x_ub, t_ub, col_weights, u_weights, tf_iter, newton_iter, lmbda, eps):with tf.device('/GPU:0'):batch_sz = N_f # Can adjust batch size for collocation points, here we set it to N_fn_batches =  N_f // batch_szadam_losses = []#create optimizer s for the network weights, collocation point mask, and initial boundary masktf_optimizer = tf.keras.optimizers.Adam(learning_rate = lr_adam_sa, beta_1=.99)tf_optimizer_weights = tf.keras.optimizers.Adam(learning_rate = lr_adam_sa, beta_1=.99)tf_optimizer_u = tf.keras.optimizers.Adam(learning_rate = lr_adam_sa, beta_1=.99)print("Starting ADAM training")# For mini-batch (if used)for epoch in range(tf_iter):for i in range(n_batches):x0_batch = x0t0_batch = t0u0_batch = u0x_f_batch = x_f[i*batch_sz:(i*batch_sz + batch_sz),]t_f_batch = t_f[i*batch_sz:(i*batch_sz + batch_sz),]loss_value, mse_0, mse_b, mse_f, grads, grads_col, grads_u, g_u, g_f = grad(u_model, x_f_batch, t_f_batch, x0_batch, t0_batch,  u0_batch, x_lb, t_lb, x_ub, t_ub, col_weights, u_weights, lmbda, eps)tf_optimizer.apply_gradients(zip(grads, u_model.trainable_variables))tf_optimizer_weights.apply_gradients(zip([-grads_col, -grads_u], [col_weights, u_weights]))if (epoch % 250 == 0) or (epoch == (tf_iter-1)):adam_losses.append(loss_value)if (epoch % 1000 == 0) or (epoch == (tf_iter-1)):print("Epoch: %d | " % (epoch), end='')tf.print(f"mse_0: {mse_0} | mse_b: {mse_b} | mse_f: {mse_f} | Total Loss: {loss_value}\n")print("Starting L-BFGS training")loss_and_flat_grad = get_loss_and_flat_grad(x_f_batch, t_f_batch, x0_batch, t0_batch, u0_batch, x_lb, t_lb, x_ub, t_ub, col_weights, u_weights, lmbda, eps)lbfgs_losses = lbfgs(loss_and_flat_grad,get_weights(u_model),Struct(), maxIter=newton_iter, learningRate=lr_lbfgs_sa)[3]return adam_losses, lbfgs_losses#L-BFGS implementation from https://github.com/pierremtb/PINNs-TF2.0
def get_loss_and_flat_grad(x_f_batch, t_f_batch, x0_batch, t0_batch, u0_batch, x_lb, t_lb, x_ub, t_ub, col_weights, u_weights, lmbda, eps):def loss_and_flat_grad(w):with tf.device('/GPU:0'):with tf.GradientTape() as tape:set_weights(u_model, w, sizes_w, sizes_b)loss_value, _, _, _ = loss(x_f_batch, t_f_batch, x0_batch, t0_batch, u0_batch, x_lb, t_lb, x_ub, t_ub, col_weights, u_weights, lmbda, eps)grad = tape.gradient(loss_value, u_model.trainable_variables)grad_flat = []for g in grad:grad_flat.append(tf.reshape(g, [-1]))grad_flat = tf.concat(grad_flat, 0)return loss_value, grad_flatreturn loss_and_flat_graddef predict(X_star):with tf.device('/GPU:0'):X_star = tf.convert_to_tensor(X_star, dtype=tf.float32)u_star, _ = u_x_model(u_model, X_star[:,0:1], X_star[:,1:2])return u_star.numpy()sizes_w = []
sizes_b = []
with tf.device('/GPU:0'):for q, width in enumerate(layers_sa):if q != 1:sizes_w.append(int(width * layers_sa[1]))sizes_b.append(int(width if q != 0 else layers_sa[1]))###############################################################################
################################ Training Loop ################################
###############################################################################
for i in range(0, number_of_neurons):print("******************************************************************")########################### SA PINN Training ############################ac_sa_train = ac_neurons[i]lmbda, eps  = float(ac_sa_train[0]), float(ac_sa_train[1])col_weights = tf.Variable(tf.random.uniform([N_f, 1]))u_weights   = tf.Variable(100*tf.random.uniform([N0, 1]))#initialize the NNu_model = neural_net(layers_sa)if (i+1 == number_of_neurons):print(f"Begin Final SA-PINN Training: lambda={lmbda}, eps={eps} (Obtaining Neuron {i+1})")else:print(f"Begin SA-PINN Training: lambda={lmbda}, eps={eps} (Obtaining Neuron {i+1})")pinn_train_time_1 = time.perf_counter()sa_losses = fit(x_f, t_f, x0, t0, u0, x_lb, t_lb, x_ub, t_ub, col_weights, u_weights, tf_iter = epochs_adam_sa, newton_iter = epochs_lbfgs_sa, lmbda=lmbda, eps=eps)pinn_train_time_2 = time.perf_counter()print("\nSA-PINN Training Completed")print(f"SA-PINN Training Time: {(pinn_train_time_2-pinn_train_time_1)/3600} Hours")w1 = u_model.layers[0].get_weights()[0].Tw2 = u_model.layers[1].get_weights()[0].Tw3 = u_model.layers[2].get_weights()[0].Tw4 = u_model.layers[3].get_weights()[0].Tw5 = u_model.layers[4].get_weights()[0].Tb1 = u_model.layers[0].get_weights()[1]b2 = u_model.layers[1].get_weights()[1]b3 = u_model.layers[2].get_weights()[1]b4 = u_model.layers[3].get_weights()[1]b5 = u_model.layers[4].get_weights()[1]P_list[i] = P(w1, w2, w3, w4, w5, b1, b2, b3, b4, b5).to(device)print(f"\nCurrent GPT-PINN Depth: [2,{i+1},1]")if (save_data):        path = fr".\Full-PINN-Data (AC)\({ac_sa_train})"if not os.path.exists(path):os.makedirs(path)np.savetxt(fr"{path}\saved_w1.txt", w1.numpy())np.savetxt(fr"{path}\saved_w2.txt", w2.numpy())np.savetxt(fr"{path}\saved_w3.txt", w3.numpy())np.savetxt(fr"{path}\saved_w4.txt", w3.numpy())np.savetxt(fr"{path}\saved_w5.txt", w3.numpy())np.savetxt(fr"{path}\saved_b1.txt", b1.numpy())np.savetxt(fr"{path}\saved_b2.txt", b2.numpy())np.savetxt(fr"{path}\saved_b3.txt", b3.numpy())np.savetxt(fr"{path}\saved_b4.txt", b3.numpy())np.savetxt(fr"{path}\saved_b5.txt", b3.numpy())x_test = xt_test[:,0].view(-1).cpu().detach().numpy()t_test = xt_test[:,1].view(-1).cpu().detach().numpy()X_star = np.hstack((x_test[:,None], t_test[:,None]))u = predict(X_star)np.savetxt(fr"{path}\u_test.txt", u)np.savetxt(fr"{path}\adam_losses.txt",  sa_losses[0])np.savetxt(fr"{path}\lbfgs_losses.txt", sa_losses[1])if (plot_pinn_sol):x_test = xt_test[:,0].view(-1).cpu().detach().numpy()t_test = xt_test[:,1].view(-1).cpu().detach().numpy()X_star = np.hstack((x_test[:,None], t_test[:,None]))u = predict(X_star)AC_plot(t_test, x_test, u, title=fr"SA-PINN Solution $\lambda={lmbda}, \epsilon={eps}$")if (plot_pinn_loss):adam_loss  = sa_losses[0]lbfgs_loss = sa_losses[1]loss_plot(epochs_adam_sa, epochs_lbfgs_sa, adam_loss, lbfgs_loss,title=fr"SA-PINN Losses $\lambda={lmbda}, \epsilon={eps}$")if (i == number_of_neurons-1) and (train_final_gpt == False):break############################ GPT-PINN Training ############################layers_gpt = np.array([2, i+1, 1])P_t, P_xx = autograd_calculations(xt_resid, P_list[i])P_t_term[:,i][:,None]  = P_tP_xx_term[:,i][:,None] = P_xxP_IC_values[:,i][:,None]    = P_list[i](IC_xt)P_BC_values[:,i][:,None]    = P_list[i](BC_xt)P_resid_values[:,i][:,None] = P_list[i](xt_resid)largest_case = 0largest_loss = 0if (i+1 == number_of_neurons):print("\nBegin Final GPT-PINN Training (Largest Loss Training)")else:print(f"\nBegin GPT-PINN Training (Finding Neuron {i+2} / Largest Loss Training)")gpt_train_time_1 = time.perf_counter()for ac_param in ac_training:lmbda, eps = ac_param[0], ac_param[1]if ([lmbda, eps] in ac_neurons):idx = ac_neurons.index([lmbda, eps])c_initial = torch.full((1,i+1), 0.)c_initial[0][idx] = 1.else:c_initial = torch.full((1,i+1), 1/(i+1))Pt_lPxx_eP_term = Pt_lPxx_eP(P_t_term[:,0:i+1], P_xx_term[:,0:i+1], P_resid_values[:,0:i+1], lmbda, eps) GPT_NN = GPT(layers_gpt, lmbda, eps, P_list[0:i+1], c_initial, xt_resid, IC_xt, BC_xt, IC_u, BC_u, f_hat, P_resid_values[:,0:i+1], P_IC_values[:,0:i+1],P_BC_values[:,0:i+1], Pt_lPxx_eP_term[:,0:i+1]).to(device)gpt_losses = gpt_train(GPT_NN, lmbda, eps, xt_resid, IC_xt, BC_xt, IC_u, BC_u,P_resid_values[:,0:i+1], P_IC_values[:,0:i+1], P_BC_values[:,0:i+1], Pt_lPxx_eP_term[:,0:i+1],lr_gpt, epochs_gpt, largest_loss, largest_case)largest_loss = gpt_losses[0]largest_case = gpt_losses[1]gpt_train_time_2 = time.perf_counter()loss_list[i] = largest_lossprint("GPT-PINN Training Completed")print(f"GPT Training Time ({i+1} Neurons): {(gpt_train_time_2-gpt_train_time_1)/3600} Hours")if (i+1 < number_of_neurons):ac_neurons[i+1] = largest_caseprint(f"\nLargest Loss (Using {i+1} Neurons): {largest_loss}")print(f"Parameter Case: {largest_case}")
total_train_time_2 = time.perf_counter()###############################################################################
# Results of largest loss, parameters chosen, and times may vary based on
# the initialization of full PINN and the final loss of the full PINN
print("******************************************************************")
print("*** Full PINN and GPT-PINN Training Complete ***")
print(f"Total Training Time: {(total_train_time_2-total_train_time_1)/3600} Hours\n")
print(f"Final GPT-PINN Depth: {[2,len(P_list),1]}")
print(f"\nActivation Function Parameters: \n{ac_neurons}\n")for j in range(number_of_neurons-1):print(f"Largest Loss of GPT-PINN Depth {[2,j+1,2]}: {loss_list[j]}")
if (train_final_gpt):print(f"Largest Loss of GPT-PINN Depth {[2,j+2,2]}: {loss_list[-1]}")if (plot_largest_loss):plt.figure(dpi=150, figsize=(10,8))if (train_final_gpt):range_end = number_of_neurons + 1list_end  = number_of_neuronselse:range_end = number_of_neurons list_end  = number_of_neurons - 1plt.plot(range(1,range_end), loss_list[:list_end], marker='o', markersize=7, c="k", linewidth=3)plt.grid(True)plt.xlim(1,max(range(1,range_end)))plt.xticks(range(1,range_end))plt.yscale("log") plt.xlabel("Number of Neurons",      fontsize=17.5)plt.ylabel("Largest Loss",           fontsize=17.5)plt.title("GPT-PINN Largest Losses", fontsize=17.5)plt.show()############################### GPT-PINN Testing ############################## 
ac_test = ac_training.tolist()
for i in ac_neurons: if (i in ac_test):ac_test.remove(i)    idx = np.random.choice(len(ac_test), test_cases, replace=False)
ac_test = np.array(ac_test)[idx]print(f"\nBegin GPT-PINN Testing ({len(set(idx.flatten()))} Cases)")I = len(P_list)   
layers_gpt = np.array([2, I, 1])
c_initial  = torch.full((1,I), 1/(I))    total_test_time_1 = time.perf_counter()
incremental_test_times = np.ones(len(ac_test))
cnt = 0for ac_test_param in ac_test:lmbda, eps = ac_test_param[0], ac_test_param[1]Pt_lPxx_eP_term = Pt_lPxx_eP(P_t_term, P_xx_term, P_resid_values, lmbda, eps) GPT_NN = GPT(layers_gpt, lmbda, eps, P_list, c_initial, xt_resid, IC_xt, BC_xt, IC_u, BC_u, f_hat, P_resid_values, P_IC_values,P_BC_values, Pt_lPxx_eP_term).to(device)gpt_losses = gpt_train(GPT_NN, lmbda, eps, xt_resid, IC_xt, BC_xt, IC_u, BC_u,P_resid_values, P_IC_values, P_BC_values, Pt_lPxx_eP_term, lr_gpt, epochs_gpt, testing=True)incremental_test_times[cnt] = (time.perf_counter()-total_test_time_1)/3600cnt += 1#np.savetxt(".\incremental_test_times.txt", incremental_test_times)total_test_time_2 = time.perf_counter()
print("\nGPT-PINN Testing Completed")
print(f"\nTotal Testing Time: {(total_test_time_2-total_test_time_1)/3600} Hours")init_time = (total_train_time_2-total_train_time_1)/3600
test_time = incremental_test_times
line = test_time + init_time
x = range(1,test_time.shape[0]+1)
plt.figure(dpi=150, figsize=(10,8))
plt.plot(x, line, c="k", lw=3.5)
plt.xlabel("Test Case Number", fontsize=22.5)
plt.ylabel("Time (Hours)", fontsize=22.5)
plt.xlim(min(x),max(x))
plt.ylim(min(line),max(line))
xtick = list(range(0,test_cases+1,5))
xtick[0] = 1
plt.xticks(xtick)
plt.grid(True)
plt.show()

结论

本论文中作者提出的GPT - PINN可解决PINN在参数化偏微分方程环境下面临的训练成本高和过度参数化两大难题。GPT - PINN是一种超精简网络，其激活函数为预训练的完整PINN，通过对三个不同的参数方程族进行测试，实验结果表明选择少量精心挑选的网络能在整个参数域准确高效地生成替代PINN。

缺点以及后续展望

GPT - PINN对参数维度敏感（高维参数仍需更多预训练模型）且对解空间快速变化（如移动激波）的适应性有限。
最后，作者提出后续可进一步探索处理参数移动时产生的不连续性，如采用非线性模型降阶策略或整合转换策略。此外，还可继续优化GPT - PINN，降低全PINN模拟次数，提高效率。最后也可尝试将该方法应用于更多类型的PDEs或其他相关领域。