Diffusion equation with training points resampling

Problem setup

We will solve a diffusion equation with training points resampling:

\[\frac{\partial y}{\partial t} = \frac{\partial^2y}{\partial x^2} - e^{-t}(\sin(\pi x) - \pi^2\sin(\pi x)), \qquad x \in [-1, 1], \quad t \in [0, 1]\]

with the initial condition

\[y(x, 0) = \sin(\pi x)\]

and the Dirichlet boundary condition

\[y(-1, t) = y(1, t) = 0.\]

The reference solution is \(y = e^{-t} \sin(\pi x)\).

Implementation

This description goes through the implementation of a solver for the above described diffusion equation step-by-step.

First, the DeepXDE, NumPy (np), and TensorFlow (tf) modules are imported:

import deepxde as dde
import numpy as np
from deepxde.backend import tf

We begin by defining computational geometries. We can use a built-in class Interval and TimeDomain and we combine both the domains using GeometryXTime as follows

geom = dde.geometry.Interval(-1, 1)
timedomain = dde.geometry.TimeDomain(0, 1)
geomtime = dde.geometry.GeometryXTime(geom, timedomain)

Next, we express the PDE residual of the diffusion equation:

def pde(x, y):
    dy_t = dde.grad.jacobian(y, x, j=1)
    dy_xx = dde.grad.hessian(y, x, j=0)
    return (
        dy_t
        - dy_xx
        + tf.exp(-x[:, 1:]
        * (tf.sin(np.pi * x[:, 0:1]) - np.pi ** 2 * tf.sin(np.pi * x[:, 0:1]))
     )

The first argument to pde is 2-dimensional vector where the first component(x[:,0:1]) is \(x\)-coordinate and the second component (x[:,1:]) is the \(t\)-coordinate. The second argument is the network output, i.e., the solution \(y(x, t)\).

Next, we consider the boundary/initial condition. on_boundary is chosen here to use the whole boundary of the computational domain as the boundary condition. We include the geotime space , time geometry created above and on_boundary as the BC in the DirichletBC function of DeepXDE. We also define IC which is the initial condition for the diffusion equation and we use the computational domain, initial function, and on_initial to specify the IC.

bc = dde.icbc.DirichletBC(geomtime, func, lambda _, on_boundary: on_boundary)
ic = dde.icbc.IC(geomtime, func, lambda _, on_initial: on_initial)

The reference solution func is defined as:

def func(x):
    return np.sin(np.pi * x[:, 0:1]) * np.exp(-x[:, 1:])

Now, we have specified the geometry, the PDE residual and the boundary/initial conditions. We then define the TimePDE problem as

data = dde.data.TimePDE(
    geomtime,
    pde,
    [bc, ic],
    num_domain=40,
    num_boundary=20,
    num_initial=10,
    train_distribution="pseudo",
    solution=func,
    num_test=10000,
)

The number 40 is the number of training residual points sampled inside the domain, and the number 20 is the number of training points sampled on the boundary (the left and right endpoints of the interval). We also include 10 initial residual points for the initial conditions and 10000 points for testing the PDE residual. The argument train_distribution="pseudo" means that the sample training points follows a pseudo-random distribution.

Next, we choose the network. Here, we use a fully connected neural network of depth 4 (i.e., 3 hidden layers) and width 32:

layer_size = [2] + [32] * 3 + [1]
activation = "tanh"
initializer = "Glorot uniform"
net = dde.nn.FNN(layer_size, activation, initializer)

The following code is to apply mini-batch gradient descent sampling method. The period is the period of resamping. Here, the training points in the domain will be resampled every 100 iterations.

resampler = dde.callbacks.PDEPointResampler(period=100)

Now, we have the PDE problem and the network. We build a Model and choose the optimizer and learning rate. We then train the model for 2000 iterations.

model = dde.Model(data, net)
model.compile("adam", lr=0.001, metrics=["l2 relative error"])
losshistory, train_state = model.train(iterations=2000, callbacks=[resampler])

We also save and plot the best trained result and loss history.

dde.saveplot(losshistory, train_state, issave=True, isplot=True)

Complete code

"""Backend supported: tensorflow.compat.v1, tensorflow, pytorch, jax, paddle"""
import deepxde as dde
import numpy as np
# Backend tensorflow.compat.v1 or tensorflow
from deepxde.backend import tf
# Backend pytorch
# import torch
# Backend jax
# import jax.numpy as jnp
# Backend paddle
# import paddle


def pde(x, y):
    # Most backends
    dy_t = dde.grad.jacobian(y, x, i=0, j=1)
    dy_xx = dde.grad.hessian(y, x, i=0, j=0)
    # Backend jax
    # dy_t, _ = dde.grad.jacobian(y, x, i=0, j=1)
    # dy_xx, _ = dde.grad.hessian(y, x, i=0, j=0)
    # Backend tensorflow.compat.v1 or tensorflow
    return (
        dy_t
        - dy_xx
        + tf.exp(-x[:, 1:])
        * (tf.sin(np.pi * x[:, 0:1]) - np.pi ** 2 * tf.sin(np.pi * x[:, 0:1]))
    )
    # Backend pytorch
    # return (
    #     dy_t
    #     - dy_xx
    #     + torch.exp(-x[:, 1:])
    #     * (torch.sin(np.pi * x[:, 0:1]) - np.pi ** 2 * torch.sin(np.pi * x[:, 0:1]))
    # )
    # Backend jax
    # return (
    #     dy_t
    #     - dy_xx
    #     + jnp.exp(-x[:, 1:])
    #     * (jnp.sin(np.pi * x[..., 0:1]) - np.pi ** 2 * jnp.sin(np.pi * x[..., 0:1]))
    # )
    # Backend paddle
    # return (
    #     dy_t
    #     - dy_xx
    #     + paddle.exp(-x[:, 1:])
    #     * (paddle.sin(np.pi * x[:, 0:1]) - np.pi ** 2 * paddle.sin(np.pi * x[:, 0:1]))
    # )


def func(x):
    return np.sin(np.pi * x[:, 0:1]) * np.exp(-x[:, 1:])


geom = dde.geometry.Interval(-1, 1)
timedomain = dde.geometry.TimeDomain(0, 1)
geomtime = dde.geometry.GeometryXTime(geom, timedomain)

bc = dde.icbc.DirichletBC(geomtime, func, lambda _, on_boundary: on_boundary)
ic = dde.icbc.IC(geomtime, func, lambda _, on_initial: on_initial)
data = dde.data.TimePDE(
    geomtime,
    pde,
    [bc, ic],
    num_domain=40,
    num_boundary=20,
    num_initial=10,
    train_distribution="pseudo",
    solution=func,
    num_test=10000,
)

layer_size = [2] + [32] * 3 + [1]
activation = "tanh"
initializer = "Glorot uniform"
net = dde.nn.FNN(layer_size, activation, initializer)

model = dde.Model(data, net)

resampler = dde.callbacks.PDEPointResampler(period=100)
model.compile("adam", lr=0.001, metrics=["l2 relative error"])
losshistory, train_state = model.train(iterations=2000, callbacks=[resampler])

dde.saveplot(losshistory, train_state, issave=True, isplot=True)