基于FNO求解一维Burgers
概述
计算流体力学是21世纪流体力学领域的重要技术之一,其通过使用数值方法在计算机中对流体力学的控制方程进行求解,从而实现流动的分析、预测和控制。传统的有限元法(finite element method,FEM)和有限差分法(finite difference method,FDM)常用于复杂的仿真流程(物理建模、网格划分、数值离散、迭代求解等)和较高的计算成本,往往效率低下。因此,借助AI提升流体仿真效率是十分必要的。
近年来,随着神经网络的迅猛发展,为科学计算提供了新的范式。经典的神经网络是在有限维度的空间进行映射,只能学习与特定离散化相关的解。与经典神经网络不同,傅里叶神经算子(Fourier Neural Operator,FNO)是一种能够学习无限维函数空间映射的新型深度学习架构。该架构可直接学习从任意函数参数到解的映射,用于解决一类偏微分方程的求解问题,具有更强的泛化能力。更多信息可参考Fourier Neural Operator for Parametric Partial Differential Equations。
本案例教程介绍利用傅里叶神经算子的1-d Burgers方程求解方法。
伯格斯方程(Burgers’ equation)
一维伯格斯方程(1-d Burgers’ equation)是一个非线性偏微分方程,具有广泛应用,包括一维粘性流体流动建模。它的形式如下:
其中\(u\)表示速度场,\(u_0\)表示初始条件,\(\nu\)表示粘度系数。
问题描述
本案例利用Fourier Neural Operator学习初始状态到下一时刻状态的映射,实现一维Burgers’方程的求解:
技术路径
MindFlow求解该问题的具体流程如下:
创建数据集。
构建模型。
优化器与损失函数。
定义求解器。
定义回调函数。
模型训练。
Fourier Neural Operator
Fourier Neural Operator模型构架如下图所示。图中\(w_0(x)\)表示初始涡度,通过Lifting Layer实现输入向量的高维映射,然后将映射结果作为Fourier Layer的输入,进行频域信息的非线性变换,最后由Decoding Layer将变换结果映射至最终的预测结果\(w_1(x)\)。
Lifting Layer、Fourier Layer以及Decoding Layer共同组成了Fourier Neural Operator。
Fourier Layer网络结构如下图所示。图中V表示输入向量,上框表示向量经过傅里叶变换后,经过线性变换R,过滤高频信息,然后进行傅里叶逆变换;另一分支经过线性变换W,最后通过激活函数,得到Fourier Layer输出向量。
[1]:
import os
import numpy as np
from mindspore import context, nn, Tensor, set_seed
from mindspore import DynamicLossScaleManager, LossMonitor, TimeMonitor
下述src
包可以在applications/data_driven/burgers/src下载。
[ ]:
from mindflow import FNO1D, RelativeRMSELoss, Solver, load_yaml_config, get_warmup_cosine_annealing_lr
from src import PredictCallback, create_training_dataset
set_seed(0)
np.random.seed(0)
[3]:
context.set_context(mode=context.GRAPH_MODE, device_target='GPU', device_id=4)
[4]:
config = load_yaml_config("burgers1d.yaml")
data_params = config["data"]
model_params = config["model"]
optimizer_params = config["optimizer"]
callback_params = config["callback"]
创建数据集
下载训练与测试数据集: data_driven/burgers/dataset。
本案例根据Zongyi Li在 Fourier Neural Operator for Parametric Partial Differential Equations 一文中对数据集的设置生成训练数据集与测试数据集。具体设置如下: 基于周期性边界,生成满足如下分布的初始条件\(u_0(x)\):
本案例选取粘度系数\(\nu=0.1\),并使用分步法求解方程,其中热方程部分在傅里叶空间中精确求解,然后使用前向欧拉方法求解非线性部分。训练集样本量为1000个,测试集样本量为200个。
[5]:
# create training dataset
train_dataset = create_training_dataset(data_params, shuffle=True)
# create test dataset
test_input, test_label = np.load(os.path.join(data_params["path"], "test/inputs.npy")), \
np.load(os.path.join(data_params["path"], "test/label.npy"))
Data preparation finished
input_path: (1000, 1024, 1)
label_path: (1000, 1024)
构建模型
网络由1层Lifting layer、1层Decoding layer以及多层Fourier Layer叠加组成:
Lifting layer对应样例代码中
FNO1D.fc0
,将输出数据\(x\)映射至高维;多层Fourier Layer的叠加对应样例代码中
FNO1D.fno_seq
,本案例采用离散傅里叶变换实现时域与频域的转换;Decoding layer对应代码中
FNO1D.fc1
与FNO1D.fc2
,获得最终的预测值。
基于上述网络结构,进行模型初始化,其中模型参数可在配置文件中修改。
[6]:
model = FNO1D(in_channels=model_params["in_channels"],
out_channels=model_params["out_channels"],
resolution=model_params["resolution"],
modes=model_params["modes"],
channels=model_params["width"],
depth=model_params["depth"])
优化器与损失函数
使用相对均方根误差作为网络训练损失函数:
[7]:
steps_per_epoch = train_dataset.get_dataset_size()
lr = get_warmup_cosine_annealing_lr(lr_init=optimizer_params["initial_lr"],
last_epoch=optimizer_params["train_epochs"],
steps_per_epoch=steps_per_epoch,
warmup_epochs=1)
optimizer = nn.Adam(model.trainable_params(), learning_rate=Tensor(lr))
loss_scale = DynamicLossScaleManager()
loss_fn = RelativeRMSELoss()
定义求解器
Solver类是模型训练和推理的接口。输入优化器、网络模型、损失函数、损失缩放策略等,即可定义求解器对象solver。代码中optimizer_params、model_params对应各项参数均在配置文件中修改。
[8]:
solver = Solver(model,
optimizer=optimizer,
loss_scale_manager=loss_scale,
loss_fn=loss_fn,
)
定义回调函数
[9]:
summary_dir = os.path.join(callback_params["summary_dir"], "FNO1D")
print(summary_dir)
pred_cb = PredictCallback(model=model,
inputs=test_input,
label=test_label,
config=config,
summary_dir=summary_dir)
./FNO1D
check test dataset shape: (200, 1024, 1), (200, 1024)
模型训练
调用求解器接口进行模型训练,调用回调接口进行评估。
[10]:
solver.train(epoch=optimizer_params["train_epochs"],
train_dataset=train_dataset,
callbacks=[LossMonitor(), TimeMonitor(), pred_cb],
dataset_sink_mode=True)
epoch: 1 step: 125, loss is 2.377823
Train epoch time: 5782.938 ms, per step time: 46.264 ms
epoch: 2 step: 125, loss is 0.88470775
Train epoch time: 1150.446 ms, per step time: 9.204 ms
epoch: 3 step: 125, loss is 0.98071647
Train epoch time: 1135.464 ms, per step time: 9.084 ms
epoch: 4 step: 125, loss is 0.5404751
Train epoch time: 1114.245 ms, per step time: 8.914 ms
epoch: 5 step: 125, loss is 0.39976493
Train epoch time: 1125.107 ms, per step time: 9.001 ms
epoch: 6 step: 125, loss is 0.508416
Train epoch time: 1127.477 ms, per step time: 9.020 ms
epoch: 7 step: 125, loss is 0.42839915
Train epoch time: 1125.775 ms, per step time: 9.006 ms
epoch: 8 step: 125, loss is 0.28270185
Train epoch time: 1118.428 ms, per step time: 8.947 ms
epoch: 9 step: 125, loss is 0.24137405
Train epoch time: 1121.705 ms, per step time: 8.974 ms
epoch: 10 step: 125, loss is 0.22623646
Train epoch time: 1118.699 ms, per step time: 8.950 ms
================================Start Evaluation================================
mean rms_error: 0.03270653011277318
=================================End Evaluation=================================
...
predict total time: 0.5012176036834717 s
epoch: 91 step: 125, loss is 0.026378194
Train epoch time: 1119.095 ms, per step time: 8.953 ms
epoch: 92 step: 125, loss is 0.057838168
Train epoch time: 1116.712 ms, per step time: 8.934 ms
epoch: 93 step: 125, loss is 0.034773324
Train epoch time: 1107.931 ms, per step time: 8.863 ms
epoch: 94 step: 125, loss is 0.029720988
Train epoch time: 1109.336 ms, per step time: 8.875 ms
epoch: 95 step: 125, loss is 0.02933883
Train epoch time: 1111.804 ms, per step time: 8.894 ms
epoch: 96 step: 125, loss is 0.03140598
Train epoch time: 1116.788 ms, per step time: 8.934 ms
epoch: 97 step: 125, loss is 0.03695058
Train epoch time: 1115.020 ms, per step time: 8.920 ms
epoch: 98 step: 125, loss is 0.039841708
Train epoch time: 1120.316 ms, per step time: 8.963 ms
epoch: 99 step: 125, loss is 0.039001673
Train epoch time: 1134.618 ms, per step time: 9.077 ms
epoch: 100 step: 125, loss is 0.038434036
Train epoch time: 1116.549 ms, per step time: 8.932 ms
================================Start Evaluation================================
mean rms_error: 0.005707952339434996
=================================End Evaluation=================================
predict total time: 0.5055065155029297 s