Hide code cell source
import matplotlib.pyplot as plt
%matplotlib inline
import matplotlib_inline
matplotlib_inline.backend_inline.set_matplotlib_formats('svg')
import seaborn as sns
sns.set_context("paper")
sns.set_style("ticks");

!pip show equinox || echo equinox not found. Installing... && pip install equinox 2> /dev/null
Hide code cell output
Name: equinox
Version: 0.11.2
Summary: Elegant easy-to-use neural networks in JAX.
Home-page: 
Author: 
Author-email: Patrick Kidger <contact@kidger.site>
License: Apache License
                           Version 2.0, January 2004
                        http://www.apache.org/licenses/

   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION

   1. Definitions.

      "License" shall mean the terms and conditions for use, reproduction,
      and distribution as defined by Sections 1 through 9 of this document.

      "Licensor" shall mean the copyright owner or entity authorized by
      the copyright owner that is granting the License.

      "Legal Entity" shall mean the union of the acting entity and all
      other entities that control, are controlled by, or are under common
      control with that entity. For the purposes of this definition,
      "control" means (i) the power, direct or indirect, to cause the
      direction or management of such entity, whether by contract or
      otherwise, or (ii) ownership of fifty percent (50%) or more of the
      outstanding shares, or (iii) beneficial ownership of such entity.

      "You" (or "Your") shall mean an individual or Legal Entity
      exercising permissions granted by this License.

      "Source" form shall mean the preferred form for making modifications,
      including but not limited to software source code, documentation
      source, and configuration files.

      "Object" form shall mean any form resulting from mechanical
      transformation or translation of a Source form, including but
      not limited to compiled object code, generated documentation,
      and conversions to other media types.

      "Work" shall mean the work of authorship, whether in Source or
      Object form, made available under the License, as indicated by a
      copyright notice that is included in or attached to the work
      (an example is provided in the Appendix below).

      "Derivative Works" shall mean any work, whether in Source or Object
      form, that is based on (or derived from) the Work and for which the
      editorial revisions, annotations, elaborations, or other modifications
      represent, as a whole, an original work of authorship. For the purposes
      of this License, Derivative Works shall not include works that remain
      separable from, or merely link (or bind by name) to the interfaces of,
      the Work and Derivative Works thereof.

      "Contribution" shall mean any work of authorship, including
      the original version of the Work and any modifications or additions
      to that Work or Derivative Works thereof, that is intentionally
      submitted to Licensor for inclusion in the Work by the copyright owner
      or by an individual or Legal Entity authorized to submit on behalf of
      the copyright owner. For the purposes of this definition, "submitted"
      means any form of electronic, verbal, or written communication sent
      to the Licensor or its representatives, including but not limited to
      communication on electronic mailing lists, source code control systems,
      and issue tracking systems that are managed by, or on behalf of, the
      Licensor for the purpose of discussing and improving the Work, but
      excluding communication that is conspicuously marked or otherwise
      designated in writing by the copyright owner as "Not a Contribution."

      "Contributor" shall mean Licensor and any individual or Legal Entity
      on behalf of whom a Contribution has been received by Licensor and
      subsequently incorporated within the Work.

   2. Grant of Copyright License. Subject to the terms and conditions of
      this License, each Contributor hereby grants to You a perpetual,
      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
      copyright license to reproduce, prepare Derivative Works of,
      publicly display, publicly perform, sublicense, and distribute the
      Work and such Derivative Works in Source or Object form.

   3. Grant of Patent License. Subject to the terms and conditions of
      this License, each Contributor hereby grants to You a perpetual,
      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
      (except as stated in this section) patent license to make, have made,
      use, offer to sell, sell, import, and otherwise transfer the Work,
      where such license applies only to those patent claims licensable
      by such Contributor that are necessarily infringed by their
      Contribution(s) alone or by combination of their Contribution(s)
      with the Work to which such Contribution(s) was submitted. If You
      institute patent litigation against any entity (including a
      cross-claim or counterclaim in a lawsuit) alleging that the Work
      or a Contribution incorporated within the Work constitutes direct
      or contributory patent infringement, then any patent licenses
      granted to You under this License for that Work shall terminate
      as of the date such litigation is filed.

   4. Redistribution. You may reproduce and distribute copies of the
      Work or Derivative Works thereof in any medium, with or without
      modifications, and in Source or Object form, provided that You
      meet the following conditions:

      (a) You must give any other recipients of the Work or
          Derivative Works a copy of this License; and

      (b) You must cause any modified files to carry prominent notices
          stating that You changed the files; and

      (c) You must retain, in the Source form of any Derivative Works
          that You distribute, all copyright, patent, trademark, and
          attribution notices from the Source form of the Work,
          excluding those notices that do not pertain to any part of
          the Derivative Works; and

      (d) If the Work includes a "NOTICE" text file as part of its
          distribution, then any Derivative Works that You distribute must
          include a readable copy of the attribution notices contained
          within such NOTICE file, excluding those notices that do not
          pertain to any part of the Derivative Works, in at least one
          of the following places: within a NOTICE text file distributed
          as part of the Derivative Works; within the Source form or
          documentation, if provided along with the Derivative Works; or,
          within a display generated by the Derivative Works, if and
          wherever such third-party notices normally appear. The contents
          of the NOTICE file are for informational purposes only and
          do not modify the License. You may add Your own attribution
          notices within Derivative Works that You distribute, alongside
          or as an addendum to the NOTICE text from the Work, provided
          that such additional attribution notices cannot be construed
          as modifying the License.

      You may add Your own copyright statement to Your modifications and
      may provide additional or different license terms and conditions
      for use, reproduction, or distribution of Your modifications, or
      for any such Derivative Works as a whole, provided Your use,
      reproduction, and distribution of the Work otherwise complies with
      the conditions stated in this License.

   5. Submission of Contributions. Unless You explicitly state otherwise,
      any Contribution intentionally submitted for inclusion in the Work
      by You to the Licensor shall be under the terms and conditions of
      this License, without any additional terms or conditions.
      Notwithstanding the above, nothing herein shall supersede or modify
      the terms of any separate license agreement you may have executed
      with Licensor regarding such Contributions.

   6. Trademarks. This License does not grant permission to use the trade
      names, trademarks, service marks, or product names of the Licensor,
      except as required for reasonable and customary use in describing the
      origin of the Work and reproducing the content of the NOTICE file.

   7. Disclaimer of Warranty. Unless required by applicable law or
      agreed to in writing, Licensor provides the Work (and each
      Contributor provides its Contributions) on an "AS IS" BASIS,
      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
      implied, including, without limitation, any warranties or conditions
      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
      PARTICULAR PURPOSE. You are solely responsible for determining the
      appropriateness of using or redistributing the Work and assume any
      risks associated with Your exercise of permissions under this License.

   8. Limitation of Liability. In no event and under no legal theory,
      whether in tort (including negligence), contract, or otherwise,
      unless required by applicable law (such as deliberate and grossly
      negligent acts) or agreed to in writing, shall any Contributor be
      liable to You for damages, including any direct, indirect, special,
      incidental, or consequential damages of any character arising as a
      result of this License or out of the use or inability to use the
      Work (including but not limited to damages for loss of goodwill,
      work stoppage, computer failure or malfunction, or any and all
      other commercial damages or losses), even if such Contributor
      has been advised of the possibility of such damages.

   9. Accepting Warranty or Additional Liability. While redistributing
      the Work or Derivative Works thereof, You may choose to offer,
      and charge a fee for, acceptance of support, warranty, indemnity,
      or other liability obligations and/or rights consistent with this
      License. However, in accepting such obligations, You may act only
      on Your own behalf and on Your sole responsibility, not on behalf
      of any other Contributor, and only if You agree to indemnify,
      defend, and hold each Contributor harmless for any liability
      incurred by, or claims asserted against, such Contributor by reason
      of your accepting any such warranty or additional liability.

   END OF TERMS AND CONDITIONS

   APPENDIX: How to apply the Apache License to your work.

      To apply the Apache License to your work, attach the following
      boilerplate notice, with the fields enclosed by brackets "[]"
      replaced with your own identifying information. (Don't include
      the brackets!)  The text should be enclosed in the appropriate
      comment syntax for the file format. We also recommend that a
      file or class name and description of purpose be included on the
      same "printed page" as the copyright notice for easier
      identification within third-party archives.

   Copyright [yyyy] [name of copyright owner]

   Licensed under the Apache License, Version 2.0 (the "License");
   you may not use this file except in compliance with the License.
   You may obtain a copy of the License at

       http://www.apache.org/licenses/LICENSE-2.0

   Unless required by applicable law or agreed to in writing, software
   distributed under the License is distributed on an "AS IS" BASIS,
   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
   See the License for the specific language governing permissions and
   limitations under the License.
Location: /Users/ibilion/.pyenv/versions/3.11.6/lib/python3.11/site-packages
Requires: jax, jaxtyping, typing-extensions
Required-by: diffrax, lineax, optimistix, orthojax
Requirement already satisfied: equinox in /Users/ibilion/.pyenv/versions/3.11.6/lib/python3.11/site-packages (0.11.2)
Requirement already satisfied: jax>=0.4.13 in /Users/ibilion/.pyenv/versions/3.11.6/lib/python3.11/site-packages (from equinox) (0.4.19)
Requirement already satisfied: jaxtyping>=0.2.20 in /Users/ibilion/.pyenv/versions/3.11.6/lib/python3.11/site-packages (from equinox) (0.2.25)
Requirement already satisfied: typing-extensions>=4.5.0 in /Users/ibilion/.pyenv/versions/3.11.6/lib/python3.11/site-packages (from equinox) (4.8.0)
Requirement already satisfied: ml-dtypes>=0.2.0 in /Users/ibilion/.pyenv/versions/3.11.6/lib/python3.11/site-packages (from jax>=0.4.13->equinox) (0.3.1)
Requirement already satisfied: numpy>=1.22 in /Users/ibilion/.pyenv/versions/3.11.6/lib/python3.11/site-packages (from jax>=0.4.13->equinox) (1.25.2)
Requirement already satisfied: opt-einsum in /Users/ibilion/.pyenv/versions/3.11.6/lib/python3.11/site-packages (from jax>=0.4.13->equinox) (3.3.0)
Requirement already satisfied: scipy>=1.9 in /Users/ibilion/.pyenv/versions/3.11.6/lib/python3.11/site-packages (from jax>=0.4.13->equinox) (1.11.3)
Requirement already satisfied: typeguard<3,>=2.13.3 in /Users/ibilion/.pyenv/versions/3.11.6/lib/python3.11/site-packages (from jaxtyping>=0.2.20->equinox) (2.13.3)

Physics-Informed Neural Networks (PINNs) - Forward Problems#

We can use PINNs to solve forward problems in place of traditional numerical methods. This is not recommended, as PINNs are not yet as efficient as traditional numerical methods. But, it is a good way to learn how to use PINNs and to understand their limitations. Throughout this section we will be following closely the methodology of Wang et al. (2023).

The toy problem - Steady-state heat equation#

We are going to solve a Poisson’s equation with a source term:

\[\begin{split} \begin{aligned} -k\Delta u(x, y) &= f(x, y) \quad \text{in} \quad \Omega = [0, L_x]\times [0, L_y], \\ u(x, y) &= 0 \quad \text{on} \quad \partial \Omega. \end{aligned} \end{split}\]

Making an exact solution#

We will use a common trick to construct an exact solution. We will use the following function:

\[ u(x, y) = u_0 \sin^2(\pi x/L_x) \sin^2(10\pi y/L_y). \]

The boundary conditions are satisfied by construction. The source term is:

Hide code cell source
import sympy

xs, ys, ks, u0s, Lxs, Lys = sympy.symbols('x y k u_0 L_x L_y')

us = u0s * sympy.sin(sympy.pi * xs / Lxs) ** 2 * sympy.sin(sympy.pi * ys / Lys) ** 2
fs = -ks * (sympy.diff(us, xs, 2) + sympy.diff(us, ys, 2))
sympy.simplify(fs)
\[\displaystyle \frac{2 \pi^{2} k u_{0} \left(- L_{x}^{2} \sin^{2}{\left(\frac{\pi x}{L_{x}} \right)} \cos{\left(\frac{2 \pi y}{L_{y}} \right)} - L_{y}^{2} \sin^{2}{\left(\frac{\pi y}{L_{y}} \right)} \cos{\left(\frac{2 \pi x}{L_{x}} \right)}\right)}{L_{x}^{2} L_{y}^{2}}\]

This is a common trick for constructing exact solutions. Remember it!

Let’s use the following parameters:

u0 = 500 # degrees Kelvin
k = 10.0 # thermal conductivity in W/mK
Lx = 0.1 # meters
Ly = 1.0 # meters

Enforcing the boundary conditions#

We are going to solve this boundary value problem using PINNS. Our model will be:

\[ u(x,y) = x(1-x)y(1-y)\text{MLP}(x, y), \]

where \(\text{MLP}(x, y)\) is a multi-layer perceptron, and \(x(1-x)y(1-y)\) is a polynomial that satisfies the boundary conditions. This is also a common trick for enforcing boundary conditions in PINNS.

Multi-layer perceptron#

We will be using a simple multi-layer perceptron to represent the solution of a PDE. The architecture of the network is given by:

\[ \text{MLP}(\mathbf{x}) = z^{(L)}, \]

where

\[ z^{(0)} = \mathbf{x} = (x,y), \quad z^{(l)} = g^{(l)}(W^{(l)}z^{(l-1)} + b^{(l)}), \quad l=1,\ldots,L-1, \]

The terms \(W^{(l)}\) and \(b^{(l)}\) are the weights and biases of the \(l\)-th layer, and \(g^{(l)}\) is the activation function of the \(l\)-th layer. The parameters \(\theta\) are the weights and biases of the network:

\[ \theta = \left\{\left(W^{(l)}, b^{(l)}\right)\right\}_{l=1}^L. \]

For PINNS applications, it is recommended that:

  • We use the tanh activation function all layers.

  • We use 128 to 512 neurons per layer.

  • We use 3 to 5 layers.

  • Use the Glorot initialization (this is the default in equinox).

Let’s make the model:

import equinox as eqx
import jax.numpy as jnp
import jax.random as jrandom

key = jrandom.PRNGKey(0)
key, subkey = jrandom.split(key)
# MLP parameters
width_size = 128
depth = 4
mlp = eqx.nn.MLP(2, 1, width_size, depth, jnp.tanh, key=subkey)

# This is the parameterization of the solution that satisfies the boundary conditions
u_hat = lambda x, y, mlp: x * (Lx - x) * y * (Ly - y) * mlp(jnp.array([x, y]))[0]

Let’s see how it looks like before we train it:

Hide code cell source
from jax import vmap

v_u_hat = eqx.filter_jit(vmap(u_hat, in_axes=(0, 0, None)))

x = jnp.linspace(0, Lx, 100)
y = jnp.linspace(0, Ly, 100)
X, Y = jnp.meshgrid(x, y)
u_pred = v_u_hat(X.flatten(), Y.flatten(), mlp).reshape(X.shape)

# the true solution for comparison
u_true = u0 * jnp.sin(jnp.pi * X / Lx) ** 2 * jnp.sin(jnp.pi * Y / Ly) ** 2

fig, ax = plt.subplots(1,2)
c = ax[0].contourf(X, Y, u_pred, cmap='viridis')
fig.colorbar(c, ax=ax[0])
ax[0].set_xlabel('x')
ax[0].set_ylabel('y')
ax[0].set_title('Initial point')
sns.despine(trim=True);
c = ax[1].contourf(X, Y, u_true, cmap='viridis', levels=jnp.linspace(0, 560, 10))
fig.colorbar(c, ax=ax[1])
ax[1].set_xlabel('x')
ax[1].set_ylabel('y')
ax[1].set_title('True solution')
sns.despine(trim=True)
plt.tight_layout();
../../_images/5ca77cae60422370412875c4f75537468590b542d32570ac33471754f3b02508.svg

Setting up the loss function#

We are going to train this by minimizing the following loss function:

\[ \mathcal{L}(\theta) = \int_{\Omega} \left[ \Delta u(x, y) + f(x, y) \right]^2 \, dx \, dy. \]

The first step, is to turn the loss function into an expectation. Construct the random variable \(\mathbf{X} = (X,Y)\) uniformly distributed in \(\Omega\). Then, the loss function can be written as:

\[ \mathcal{L}(\theta) = |\Omega|\mathbb{E}\left[ \left[ \Delta u(\mathbf{X}) + f(\mathbf{X}) \right]^2 \right]. \]

Notice that the volume of \(\Omega\), \(\Omega|\), appears as a constant factor. Where did it come from? Recall the probability density function of \(\mathbf{X}\) is:

\[ p(\mathbf{x}) = \frac{1}{|\Omega|} \quad \text{if} \quad \mathbf{x} \in \Omega, \quad \text{and} \quad 0 \quad \text{otherwise}. \]

So:

\[ \int_{\Omega} \left[ \Delta u(\mathbf{x}) + f(\mathbf{x}) \right]^2 \, d\mathbf{x} = \int_{\Omega} \left[ \Delta u(\mathbf{x}) + f(\mathbf{x}) \right]^2 |\Omega|p(\mathbf{x}) \, d\mathbf{x} = |\Omega|\mathbb{E}\left[ \left[ \Delta u(\mathbf{X}) + f(\mathbf{X}) \right]^2 \right]. \]

We can approximate this expectation by sampling a finite number of points from \(\Omega\) and averaging the loss function over these points. This is what we will be doing in each iteration of the training process.

from jax import grad

# First order derivatives
u_x = grad(u_hat, argnums=0)
u_y = grad(u_hat, argnums=1)

# Second order deerivatives
u_xx = grad(u_x, argnums=0)
u_yy = grad(u_y, argnums=1)

# The Laplacian
D2_u = lambda x, y, mlp: u_xx(x, y, mlp) + u_yy(x, y, mlp)

# The source term
source_term = lambda x, y: 2.0 * jnp.pi ** 2 * k * u0 * (
    -Lx ** 2 * jnp.sin(jnp.pi * x / Lx) ** 2 * jnp.cos(2.0 * jnp.pi * y / Ly)
    -Ly ** 2 * jnp.sin(jnp.pi * y / Ly) ** 2 * jnp.cos(2.0 * jnp.pi * x / Lx)
) / (Lx ** 2 * Ly ** 2)

# The PDE residual (vectorize)
pde_residual = vmap(
    lambda x, y, mlp: D2_u(x, y, mlp) + source_term(x, y),
    in_axes=(0, 0, None))

# Finally the loss function:
loss = lambda mlp, x, y: Lx * Ly * jnp.mean(jnp.square(pde_residual(x, y, mlp)))

Training the model#

We will use the ADAM optimizer with standard parameters. This is the training algorithm:

def train(
        loss,
        mlp,
        key,
        optimizer,
        Lx=1.0,
        Ly=1.0,
        num_collocation_residual=512,
        num_iter=10_000,
        freq=1,
    ):

    @eqx.filter_jit
    def step(opt_state, mlp, xs, ys):
        value, grads = eqx.filter_value_and_grad(loss)(mlp, xs, ys)
        updates, opt_state = optimizer.update(grads, opt_state)
        mlp = eqx.apply_updates(mlp, updates)
        return mlp, opt_state, value
    
    opt_state = optimizer.init(eqx.filter(mlp, eqx.is_inexact_array))
    
    losses = []
    for i in range(num_iter):
        key, subkey = jrandom.split(key)
        xb = jrandom.uniform(subkey, (num_collocation_residual,), maxval=Lx)
        key, subkey = jrandom.split(key)
        yb = jrandom.uniform(subkey, (num_collocation_residual,), maxval=Ly)
        mlp, opt_state, value = step(opt_state, mlp, xb, yb)
        if i % freq == 0:
            losses.append(value)
            print(f"Step {i}, residual loss {value:.3e}")
    return mlp, losses

And this is the actual trainning:

import optax

optimizer = optax.adam(1e-3)
trained_mlp, losses = train(loss, mlp, key, optimizer, num_collocation_residual=256, num_iter=1_000, freq=100, Lx=Lx, Ly=Ly)
Step 0, residual loss 1.694e+12
Step 100, residual loss 1.799e+12
Step 200, residual loss 1.611e+12
Step 300, residual loss 1.676e+12
Step 400, residual loss 1.715e+12
Step 500, residual loss 1.520e+12
Step 600, residual loss 1.874e+12
Step 700, residual loss 2.111e+12
Step 800, residual loss 1.660e+12
Step 900, residual loss 1.762e+12

Let’s visualize first the evolution of the loss function:

Hide code cell source
fig, ax = plt.subplots()
ax.plot(losses)
ax.set_xlabel('Iteration x 100')
ax.set_ylabel('Residual loss')
sns.despine(trim=True);
../../_images/1ce1e39b17d4d7970c48b924e717fc9d21044495950c9a95f80ce11540d946c0.svg

It’s not working! The loss function is not decreasing. Here is also the trained “solution:”

Hide code cell source
fig, ax = plt.subplots(1,2)
u_pred = v_u_hat(X.flatten(), Y.flatten(), trained_mlp).reshape(X.shape)
c = ax[0].contourf(X, Y, u_pred, cmap='viridis')
fig.colorbar(c, ax=ax[0])
ax[0].set_xlabel('x')
ax[0].set_ylabel('y')
ax[0].set_title('Trained solution')
sns.despine(trim=True);
c = ax[1].contourf(X, Y, u_true, cmap='viridis', levels=jnp.linspace(0, 560, 10))
fig.colorbar(c, ax=ax[1])
ax[1].set_xlabel('x')
ax[1].set_ylabel('y')
ax[1].set_title('True solution')
sns.despine(trim=True)
plt.tight_layout();
../../_images/7501020c1de36e2ca47cecbc25e00d452a3ddcdd779772865edc540c6d8ed16a.svg

Non-dimensionalization of Partial Differential Equations#

The problem is that the initialization of the network weights is completely off. Glorot initialization is assuming that the input data and the output data are scaled to be around zero (scale of about 1). To get our problem to that scale, we need to non-dimensionalize it.

We start by picking characteristic lengthscales. Here, it makes sense to pick \(L_x\) and \(L_y\) as characteristic lengthscales. We also need to pick a characteristic value for the temperature, say \(u_s\).

\[ \begin{aligned} \tilde{x} = \frac{x}{L_x}, \quad \tilde{y} = \frac{y}{L_y}, \quad \tilde{u} = \frac{u}{u_s}. \end{aligned} \]

The next step is to find the partial derivatives of \(u\) with respect to \(x\) and \(y\) in terms of the non-dimensional variables:

\[ \frac{\partial u}{\partial x} = \frac{u_s}{L_x}\frac{\partial \tilde{u}}{\partial \tilde{x}}, \quad \frac{\partial u}{\partial y} = \frac{u_s}{L_y}\frac{\partial \tilde{u}}{\partial \tilde{y}}. \]

Here, we just used the chain rule.

The second derivatives are:

\[ \frac{\partial^2 u}{\partial x^2} = \frac{u_s}{L_x^2}\frac{\partial^2 \tilde{u}}{\partial \tilde{x}^2}, \quad \frac{\partial^2 u}{\partial y^2} = \frac{u_s}{L_y^2}\frac{\partial^2 \tilde{u}}{\partial \tilde{y}^2}. \]

Plugging these into the original PDE, we get:

\[ -k\left(\frac{u_s}{L_x^2}\frac{\partial^2 \tilde{u}}{\partial \tilde{x}^2} + \frac{u_s}{L_y^2}\frac{\partial^2 \tilde{u}}{\partial \tilde{y}^2}\right) = f. \]

Define now the non-dimensional source term:

\[ \tilde{f} = \frac{f}{f_s}. \]

And the non-dimensional thermal conductivity:

\[ \tilde{k}_x = \frac{ku_s}{L_x^2 f_s}, \quad \tilde{k}_y = \frac{ku_s}{L_y^2 f_s}. \]

The equation becomes:

\[ -\left(\tilde{k_x}\frac{\partial^2 \tilde{u}}{\partial \tilde{x}^2} + \tilde{k_y}\frac{\partial^2 \tilde{u}}{\partial \tilde{y}^2}\right) = \tilde{f}, \]

defined on the non-dimensional domain \(\tilde{\Omega} = [0, 1]\times [0, 1]\).

Now we need to pick \(u_s\) and \(f_s\) so that both sides of the equation are of order one. We can easily make the lef-hand side to be of order one by picking:

\[ f_s = \|f\|_{\infty} = \max_{(x, y) \in \Omega} |f(x, y)|. \]

For the left-hand side, we can pick \(u_s\) so that:

\[ \left\|-\left(\tilde{k_x}\frac{\partial^2 \tilde{u}}{\partial \tilde{x}^2} + \tilde{k_y}\frac{\partial^2 \tilde{u}}{\partial \tilde{y}^2}\right)\right\| = 1. \]

Solving for \(u_s\), we get that \(u_s\) should be of the order:

\[ u_s = \frac{f_s}{k\max\{1/L_x^2, 1/L_y^2\}\max\left\{\left\|\frac{\partial^2\tilde{u}}{\partial \tilde{x}^2}\right\|, \left\|\frac{\partial^2\tilde{u}}{\partial \tilde{y}^2}\right\|\right\}}. \]

Do not take the equality sign too seriously. All we need is that \(u_s\) is of the order of the right-hand side of the equation.

Let’s find these numbers:

# These functions convert from x to xt and y to yt and vice versa
to_x = lambda xt: xt * Lx
to_y = lambda yt: yt * Ly
to_xt = lambda x: x / Lx
to_yt = lambda y: y / Ly
Xt = to_xt(X)
Yt = to_yt(Y)

fs = jnp.abs(source_term(X.flatten(), Y.flatten())).max()

# Start a random MLP
tmlp = eqx.nn.MLP(2, 1, width_size, depth, jnp.tanh, key=subkey)
tu_hat = lambda tx, ty, tmlp: tx * (1.0 - tx) * ty * (1.0 - ty) * tmlp(jnp.array([tx, ty]))[0]
# Find the required gradients
tu_x = grad(tu_hat, argnums=0)
tu_y = grad(tu_hat, argnums=1)
tu_xx = grad(tu_x, argnums=0)
tu_yy = grad(tu_y, argnums=1)
v_tu_xx = eqx.filter_jit(vmap(u_xx, in_axes=(0, 0, None)))
v_tu_yy = eqx.filter_jit(vmap(u_yy, in_axes=(0, 0, None)))
max_u_xx = jnp.abs(v_tu_xx(Xt.flatten(), Yt.flatten(), tmlp).reshape(X.shape)).max()
max_u_yy = jnp.abs(v_tu_yy(Xt.flatten(), Yt.flatten(), tmlp).reshape(X.shape)).max()

# Calculate the scale:
us = fs / k / max(max_u_xx, max_u_yy) / max(1/Lx**2, 1/Ly**2)

print(f"Scale factor fs: {fs:.2e}")
print(f"Scale factor us: {us:.2e}")
Scale factor fs: 9.96e+06
Scale factor us: 4.40e+04

Let’s check if the left hand side is comparable to the right hand side:

lhs = lambda tx, ty, tmlp: -k * us / fs * (1.0 / Lx ** 2 * tu_xx(tx, ty, tmlp) + 1.0 / Ly ** 2 * tu_yy(tx, ty, tmlp))
v_lhs = eqx.filter_jit(vmap(lhs, in_axes=(0, 0, None)))

fig, ax = plt.subplots(1,2)
tl = v_lhs(Xt.flatten(), Yt.flatten(), tmlp).reshape(X.shape)
c = ax[0].contourf(Xt, Yt, tl, cmap='viridis')
fig.colorbar(c, ax=ax[0])
ax[0].set_xlabel(r'$\tilde{x}$')
ax[0].set_ylabel(r'$\tilde{y}$')
ax[0].set_title('Left-hand side')

tf = source_term(X.flatten(), Y.flatten()).reshape(X.shape) / fs
c = ax[1].contourf(Xt, Yt, tf, cmap='viridis')
fig.colorbar(c, ax=ax[1])
ax[1].set_xlabel(r'$\tilde{x}$')
ax[1].set_ylabel(r'$\tilde{y}$')
ax[1].set_title('Scaled source term')

sns.despine(trim=True)
plt.tight_layout();
../../_images/0593d442c98155d917189fa65538fb023dcff9f3aa873ffebe7f35745d2fe297.svg

This looks pretty good as a starting point!

Here is what the scaled conductivity looks like:

tkx = (k * us) / (Lx ** 2 * fs)
tky = (k * us) / (Ly ** 2 * fs)
print(f"tkx = {tkx:.3e}, tky = {tky:.3e}") 
tkx = 4.415e+00, tky = 4.415e-02

Let’s now try training the scaled version of the problem:

tmlp = eqx.nn.MLP(2, 1, width_size, depth, jnp.tanh, key=subkey)
tu_hat = lambda tx, ty, tmlp: tx * (1.0 - tx) * ty * (1.0 - ty) * tmlp(jnp.array([tx, ty]))[0]
tu_x = grad(tu_hat, argnums=0)
tu_y = grad(tu_hat, argnums=1)
tu_xx = grad(tu_x, argnums=0)
tu_yy = grad(tu_y, argnums=1)
tilde_source_term = lambda tx, ty: source_term(to_x(tx), to_y(ty)) / fs
tpde_residual = vmap(
    lambda tx, ty, tmlp: tkx * tu_xx(tx, ty, tmlp) + tky * tu_yy(tx, ty, tmlp) + tilde_source_term(tx, ty),
    in_axes=(0, 0, None))
tloss = lambda tmlp, tx, ty: jnp.mean(jnp.square(tpde_residual(tx, ty, tmlp)))
optimizer = optax.adam(1e-3)
trained_tmlp, losses = train(tloss, tmlp, key, optimizer, num_collocation_residual=256, num_iter=2_000, freq=100, Lx=1.0, Ly=1.0)
Step 0, residual loss 1.776e-01
Step 100, residual loss 2.152e-02
Step 200, residual loss 1.609e-02
Step 300, residual loss 9.821e-03
Step 400, residual loss 8.162e-03
Step 500, residual loss 1.964e-03
Step 600, residual loss 1.231e-03
Step 700, residual loss 3.521e-04
Step 800, residual loss 2.916e-04
Step 900, residual loss 1.718e-04
Step 1000, residual loss 2.290e-04
Step 1100, residual loss 2.774e-04
Step 1200, residual loss 1.921e-04
Step 1300, residual loss 1.431e-04
Step 1400, residual loss 1.637e-04
Step 1500, residual loss 2.281e-04
Step 1600, residual loss 1.966e-04
Step 1700, residual loss 1.857e-04
Step 1800, residual loss 1.171e-04
Step 1900, residual loss 4.257e-04

Note

If you are running this on Google Colab Pro, try using an A100 GPU. On my M1 Macbook Pro, the training takes about 17.5s. On an A100 GPU, it takes about 8s.

I save this loss for the next notebook.

import numpy as np
np.savez("mlp_losses.npz", losses=losses)

Here is the loss function:

fig, ax = plt.subplots()
ax.plot(losses)
ax.set_xlabel('Iteration x 100')
ax.set_ylabel('Residual loss')
sns.despine(trim=True);
../../_images/c8f150f76b8b287d9515a43f24e14ba7b90a0fdccdc20762b3e744c4c6ac89f1.svg

And here is the solution scaled back to the original domain:

v_tu_hat = vmap(tu_hat, in_axes=(0, 0, None))
fig, ax = plt.subplots(1,2)
u_pred = v_tu_hat(to_xt(X.flatten()), to_yt(Y.flatten()), trained_tmlp).reshape(X.shape)
c = ax[0].contourf(X, Y, u_pred * us, cmap='viridis',
    levels=jnp.linspace(-1, 510.0, 10))
fig.colorbar(c, ax=ax[0])
ax[0].set_xlabel('x')
ax[0].set_ylabel('y')
ax[0].set_title('Trained solution')
sns.despine(trim=True);
c = ax[1].contourf(X, Y, u_true, cmap='viridis',
    levels=jnp.linspace(-1, 510.0, 10))
fig.colorbar(c, ax=ax[1])
ax[1].set_xlabel('x')
ax[1].set_ylabel('y')
ax[1].set_title('Scaled true solution')
sns.despine(trim=True)
plt.tight_layout();
../../_images/c3646ee90adeab0b55a2e7807675f46d48ce4e5401b138a32483821a339f6025.svg

The point wise error is here:

fig, ax = plt.subplots()
ax.set_xlabel('x')
ax.set_ylabel('y')
ax.set_title('Trained solution')
c = ax.contourf(to_xt(X), to_yt(Y), jnp.abs(u_pred * us - u_true), cmap='viridis')
plt.colorbar(c, ax=ax)
sns.despine(trim=True);
../../_images/d18a6583b4eea3720d5955f31393c1d4120ddc858ecf2f3e83c28c18c0cd01aa.svg

It is very hard to achieve better error with plain MLPs.