## Solving coupled ODEs with a neural network and autograd

| categories: | tags: | View Comments

In a previous post I wrote about using ideas from machine learning to solve an ordinary differential equation using a neural network for the solution. A friend recently tried to apply that idea to coupled ordinary differential equations, without success. It seems like that should work, so here we diagnose the issue and figure it out. This is a long post, but it works in the end.

In the classic series reaction $$A \rightarrow B \rightarrow C$$ in a batch reactor, we get the set of coupled mole balances:

$$dC_A/dt = -k_1 C_A$$

$$dC_B/dt = k_1 C_A - k_2 C_B$$

$$dC_C/dt = k2 C_B$$

## 1 The standard numerical solution

Here is the standard numerical solution to this problem. This will give us a reference for what the solution should look like.

from scipy.integrate import solve_ivp

def ode(t, C):
Ca, Cb, Cc = C
dCbdt = k1 * Ca - k2 * Cb
dCcdt = k2 * Cb

C0 = [1.0, 0.0, 0.0]
k1 = 1
k2 = 1

sol = solve_ivp(ode, (0, 10), C0)

%matplotlib inline
import matplotlib.pyplot as plt

plt.plot(sol.t, sol.y.T)
plt.legend(['A', 'B', 'C'])
plt.xlabel('Time')
plt.ylabel('C')


## 2 Can a neural network learn the solution?

The first thing I want to show is that you can train a neural network to reproduce this solution. That is certainly a prerequisite to the idea working. We use the same code I used before, but this time our neural network will output three values, one for each concentration.

import autograd.numpy as np

def init_random_params(scale, layer_sizes, rs=npr.RandomState(0)):
"""Build a list of (weights, biases) tuples, one for each layer."""
return [(rs.randn(insize, outsize) * scale,   # weight matrix
rs.randn(outsize) * scale)           # bias vector
for insize, outsize in zip(layer_sizes[:-1], layer_sizes[1:])]

def swish(x):
"see https://arxiv.org/pdf/1710.05941.pdf"
return x / (1.0 + np.exp(-x))

def C(params, inputs):
"Neural network functions"
for W, b in params:
outputs = np.dot(inputs, W) + b
inputs = swish(outputs)
return outputs

# initial guess for the weights and biases
params = init_random_params(0.1, layer_sizes=[1, 8, 3])


Now, we train our network to reproduce the solution. I ran this block manually a bunch of times, but eventually you see that we can train a one layer network with 8 nodes to output all three concentrations pretty accurately. So, there is no issue there, a neural network can represent the solution.

def objective_soln(params, step):
return np.sum((sol.y.T - C(params, sol.t.reshape([-1, 1])))**2)

step_size=0.001, num_iters=500)

plt.plot(sol.t.reshape([-1, 1]), C(params, sol.t.reshape([-1, 1])),
sol.t, sol.y.T, 'o')
plt.legend(['A', 'B', 'C', 'Ann', 'Bnn', 'Cnn'])
plt.xlabel('Time')
plt.ylabel('C')


## 3 Given a neural network function how do we get the right derivatives?

The next issue is how do we get the relevant derivatives. The solution method I developed here relies on using optimization to find a set of weights that produces a neural network whose derivatives are consistent with the ODE equations. So, we need to be able to get the derivatives that are relevant in the equations.

The neural network outputs three concentrations, and we need the time derivatives of them. Autograd provides three options: grad, elementwise_grad and jacobian. We cannot use grad because our function is not scalar. We cannot use elementwise_grad because that will give the wrong shape (I think it may be the sum of the gradients). That leaves us with the jacobian. This, however, gives an initially unintuitive (i.e. it isn't what we need out of the box) result. The output is 4-dimensional in this case, consistent with the documentation of that function.

jacC = jacobian(C, 1)
jacC(params, sol.t.reshape([-1, 1])).shape

(17, 3, 17, 1)



Why does it have this shape? Our time input vector we used has 17 time values, in a column vector. That leads to an output from the NN with a shape of (17, 3), i.e. the concentrations of each species at each time. The jacobian will output an array of shape (17, 3, 17, 1), and we have to extract the pieces we want from that. The first and third dimensions are related to the time steps. The second dimension is the species, and the last dimension is nothing here, but is there because the input is in a column. I use some fancy indexing on the array to get the desired arrays of the derivatives. This is not obvious out of the box. I only figured this out by direct comparison of the data from a numerical solution and the output of the jacobian. Here I show how to do that, and make sure that the derivatives we pull out are comparable to the derivatives defined by the ODEs above. Parity here means they are comparable.

i = np.arange(len(sol.t))
plt.plot(jacC(params, sol.t.reshape([-1, 1]))[i, 0, i, 0],   -k1 * sol.y[0], 'ro')
plt.plot(jacC(params, sol.t.reshape([-1, 1]))[i, 1, i, 0],   -k2 * sol.y[1] + k1 * sol.y[0], 'bo')
plt.plot(jacC(params, sol.t.reshape([-1, 1]))[i, 2, i, 0],   k2 * sol.y[1], 'go')

[<matplotlib.lines.Line2D at 0x118a2e860>]



Note this is pretty inefficient. It requires a lot of calculations (the jacobian here has print(17*3*17) 867 elements) to create the jacobian, and we don't need most of them. You could avoid this by creating separate neural networks for each species, and then just use elementwise_grad on each one. Alternatively, one might be able to more efficiently compute some vector-jacobian product. Nevertheless, it looks like we can get the correct derivatives out of the neural network, we just need a convenient function to return them. Here is one such function for this problem, using a fancier slicing and reshaping to get the derivative array.

# Derivatives
jac = jacobian(C, 1)

def dCdt(params, t):
i = np.arange(len(t))
return jac(params, t)[i, :, i].reshape((len(t), 3))


## 4 Solving the system of ODEs with a neural network

Finally, we are ready to try solving the ODEs solely by the neural network approach. We reinitialize the neural network first, and define a time grid to solve it on.

t = np.linspace(0, 10, 25).reshape((-1, 1))
params = init_random_params(0.1, layer_sizes=[1, 8, 3])
i = 0    # number of training steps
N = 501  # epochs for training
et = 0.0 # total elapsed time


We define our objective function. This function will be zero at the perfect solution, and has contributions for each mole balance and the initial conditions. It could make sense to put additional penalties for things like negative concentrations, or the sum of concentrations is a constant, but we do not do that here, and it does not seem to be necessary.

def objective(params, step):
Ca, Cb, Cc = C(params, t).T
dCadt, dCbdt, dCcdt = dCdt(params, t).T

z1 = np.sum((dCadt + k1 * Ca)**2)
z2 = np.sum((dCbdt - k1 * Ca + k2 * Cb)**2)
z3 = np.sum((dCcdt - k2 * Cb)**2)
ic = np.sum((np.array([Ca[0], Cb[0], Cc[0]]) - C0)**2)  # initial conditions
return z1 + z2 + z3 + ic

def callback(params, step, g):
if step % 100 == 0:
print("Iteration {0:3d} objective {1}".format(step,
objective(params, step)))

objective(params, 0)  # make sure the objective is scalar

5.2502237371050295



Finally, we run the optimization. I also manually ran this block several times.

import time
t0 = time.time()

step_size=0.001, num_iters=N, callback=callback)

i += N
t1 = (time.time() - t0) / 60
et += t1

plt.plot(t, C(params, t), sol.t, sol.y.T, 'o')
plt.legend(['Ann', 'Bnn', 'Cnn', 'A', 'B', 'C'])
plt.xlabel('Time')
plt.ylabel('C')
print(f'{t1:1.1f} minutes elapsed this time. Total time = {et:1.2f} min. Total epochs = {i}.')

Iteration   0 objective 0.00047651643957525214
Iteration 100 objective 0.0004473301532609342
Iteration 200 objective 0.00041218410058863227
Iteration 300 objective 0.00037161526137030344
Iteration 400 objective 0.000327567400443358
Iteration 500 objective 0.0002836975879675981
0.6 minutes elapsed this time. Total time = 4.05 min. Total epochs = 3006.



The effort seems to have been worth it though, we get a pretty good solution from our neural network.

We can check the accuracy of the derivatives by noting the sum of the derivatives in this case should be zero. Here you can see that the sum is pretty small. It would take additional optimization to a lower error to get this to be smaller.

plt.plot(t, np.sum(dCdt(params, t), axis=1))
plt.xlabel('Time')
plt.ylabel(r'$\Sigma dC/dt$')


## 5 Summary

In the end, this method is illustrated to work for systems of ODEs also. There is some subtlety in how to get the relevant derivatives from the jacobian, but after that, it is essentially the same. I think it would be much faster to do this with separate neural networks for each function in the solution because then you do not need the jacobian, you can use elementwise_grad.

This is not faster than direct numerical integration. One benefit to this solution over a numerical solution is we get an actual continuous function as the solution, rather than an array of data. This solution is not reliable at longer times, but then again neither is extrapolation of numeric data. It could be interesting to explore if this has any benefits for stiff equations. Maybe another day. For now, I am declaring victory for autograd on this problem.

org-mode source

Org-mode version = 9.1.14

## A differentiable ODE integrator for sensitivity analysis

| categories: | tags: | View Comments

Last time I wrote about using automatic differentiation to find the derivative of an integral function. A related topic is finding derivatives of functions that are defined by differential equations. We typically use a numerical integrator to find solutions to these functions. Those leave us with numeric solutions which we then have to use to approximate derivatives. What if the integrator itself was differentiable? It is after all, just a program, and automatic differentiation should be able to tell us the derivatives of functions that use them. This is not a new idea, there is already a differentiable ODE solver in Tensorflow. Here I will implement a simple Runge Kutta integrator and then show how we can use automatic differentiation to do sensitivity analysis on the numeric solution.

I previously used autograd for sensitivity analysis on analytical solutions in this post. Here I will compare those results to the results from sensitivity analysis on the numerical solutions.

First, we need an autograd compatible ODE integrator. Here is one implementation of a simple, fourth order Runge-Kutta integrator. Usually, I would use indexing to do this, but that was not compatible with autograd, so I just accumulate the solution. This is a limitation of autograd, and it is probably not an issue with Tensorflow, for example, or probably pytorch. Those are more sophisticated, and more difficult to use packages than autograd. Here I am just prototyping an idea, so we stick with autograd.

import autograd.numpy as np
%matplotlib inline
import matplotlib.pyplot as plt

def rk4(f, tspan, y0, N=50):
x, h = np.linspace(*tspan, N, retstep=True)
y = []
y = y + [y0]
for i in range(0, len(x) - 1):
k1 = h * f(x[i], y[i])
k2 = h * f(x[i] + h / 2, y[i] + k1 / 2)
k3 = h * f(x[i] + h / 2, y[i] + k2 / 2)
k4 = h * f(x[i + 1], y[i] + k3)
y += [y[-1] + (k1 + (2 * k2) + (2 * k3) + k4) / 6]
return x, y


Now, we just check that it works as expected:

Ca0 = 1.0
k1 = k_1 = 3.0

def dCdt(t, Ca):
return -k1 * Ca + k_1 * (Ca0 - Ca)

t, Ca = rk4(dCdt, (0, 0.5), Ca0)

def analytical_A(t, k1, k_1):
return Ca0 / (k1 + k_1) * (k1 * np.exp(-(k1 + k_1) * t) + k_1)

plt.plot(t, Ca, label='RK4')
plt.plot(t, analytical_A(t, k1, k_1), 'r--', label='analytical')
plt.xlabel('t')
plt.ylabel('[A]')
plt.xlim([0, 0.5])
plt.ylim([0.5, 1])
plt.legend()


That looks fine, we cannot visually distinguish the two solutions, and they both look like Figure 1 in this paper. Note the analytical solution is not that complex, but it would not take much variation of the rate law to make this solution difficult to derive.

Next, to do sensitivity analysis, we need to define a function for $$A$$ that depends on the rate constants, so we can take a derivative of it with respect to the parameters we want the sensitivity from. We seek the derivatives: $$\frac{dC_A}{dk_1}$$ and $$\frac{dC_A}{dk_{-1}}$$. Here is a function that does that. It will return the value of [A] at $$t$$ given an initial concentration and the rate constants.

def A(Ca0, k1, k_1, t):
def dCdt(t, Ca):
return -k1 * Ca + k_1 * (Ca0 - Ca)
t, Ca_ = rk4(dCdt, (0, t), Ca0)
return Ca_[-1]

# Here are the two derivatives we seek.


We also use autograd to get the derivatives from the analytical solution for comparison.

dAdk1 = grad(analytical_A, 1)


Now, we can plot the sensitivities over the time range and compare them. I use the list comprehensions here because the AD functions aren't vectorized.

tspan = np.linspace(0, 0.5)

# From the numerical solutions
k1_sensitivity = [dCadk1(1.0, 3.0, 3.0, t) for t in tspan]
k_1_sensitivity = [dCadk_1(1.0, 3.0, 3.0, t) for t in tspan]

# from the analytical solutions
ak1_sensitivity = [dAdk1(t, 3.0, 3.0) for t in tspan]
ak_1_sensitivity = [dAdk_1(t, 3.0, 3.0) for t in tspan]

plt.plot(tspan, np.abs(ak1_sensitivity), 'b-', label='k1 analytical')
plt.plot(tspan, np.abs(k1_sensitivity), 'y--', label='k1 numerical')

plt.plot(tspan, np.abs(ak_1_sensitivity), 'r-', label='k_1 analytical')
plt.plot(tspan, np.abs(k_1_sensitivity), 'k--', label='k_1 numerical')

plt.xlim([0, 0.5])
plt.ylim([0, 0.1])
plt.legend()
plt.xlabel('t')
plt.ylabel('sensitivity')


The two approaches are indistinguishable on paper. I will note that it takes a lot longer to make the graph from the numerical solution than from the analytical solution because at each point you have to reintegrate the solution from the beginning, which is certainly not efficient. That is an implementation detail that could probably be solved, at the expense of making the code look different than the way I would normally think about the problem.

On the other hand, it is remarkable we get derivatives from the numerical solution, and they look really good! That means we could do sensitivity analysis on more complex reactions, and still have a reasonable way to get sensitivity. The work here is a long way from that. My simple Runge-Kutta integrator isn't directly useful for systems of ODEs, it wouldn't work well on stiff problems, the step size isn't adaptive, etc. The Tensorflow implementation might be more suitable for this though, and maybe this post is motivation to learn how to use it!

org-mode source

Org-mode version = 9.1.13

## Compressibility factor variation from the van der Waals equation by three different approaches

| categories: | tags: | View Comments

In the book Problem solving in chemical and biochemical engineering with POLYMATH, Excel and Matlab by Cutlip and Shacham there is a problem (7.1) where you want to plot the compressibility factor for CO2 over a range of $$0.1 \le P_r <= 10$$ for a constant $$T_r=1.1$$ using the van der Waal equation of state. There are a two standard ways to do this:

1. Solve a nonlinear equation for different values of $$P_r$$.
2. Solve a nonlinear equation for one value of $$P_r$$, then derive an ODE for how the compressibility varies with $$P_r$$ and integrate it over the relevant range.

In this post, we compare and contrast the two methods, and consider a variation of the second method that uses automatic differentiation.

## 1 Method 1 - fsolve

The van der Waal equation of state is:

$$P = \frac{R T}{V - b} - \frac{a}{V^2}$$.

We define the reduced pressure as $$P_r = P / P_c$$, and the reduced temperature as $$T_r = T / T_c$$.

So, we simply solve for V at a given $$P_r$$, and then compute $$Z$$. There is a subtle trick needed to make this easy to solve, and that is to multiply each side of the equation by $$(V - b)$$ to avoid a singularity when $$V = b$$, which happens in this case near $$P_r \approx 7.5$$.

from scipy.optimize import fsolve
import numpy as np
%matplotlib inline
import matplotlib.pyplot as plt

R = 0.08206
Pc = 72.9
Tc = 304.2

a = 27 * R**2 * Tc**2 / (Pc * 64)
b = R * Tc / (8 * Pc)

Tr = 1.1

def objective(V, Pr):
P = Pr * Pc
T = Tr * Tc
return P * (V - b) - (R * T)  +  a / V**2 * (V - b)

Pr_range = np.linspace(0.1, 10)
V = [fsolve(objective, 3, args=(Pr,))[0] for Pr in Pr_range]

T = Tr * Tc
P_range = Pr_range * Pc
Z = P_range * V / (R * T)

plt.plot(Pr_range, Z)
plt.xlabel('$P_r$')
plt.ylabel('Z')
plt.xlim([0, 10])
plt.ylim([0, 2])

(0, 2)



That looks like Figure 7-1 in the book. This approach is fine, but the equation did require a little algebraic finesse to solve, and you have to use some iteration to get the solution.

## 2 Method 2 - solve_ivp

In this method, you have to derive an expression for $$\frac{dV}{dP_r}$$. That derivation goes like this:

$$\frac{dV}{dP_r} = \frac{dV}{dP} \frac{dP}{dP_r}$$

The first term $$\frac{dV}{dP}$$ is $$(\frac{dP}{dV})^{-1}$$, which we can derive directly from the van der Waal equation, and the second term is just a constant: $$P_c$$ from the definition of $$P_r$$.

They derived:

$$\frac{dP}{dV} = -\frac{R T}{(V - b)^2} + \frac{2 a}{V^3}$$

We need to solve for one V, at the beginning of the range of $$P_r$$ we are interested in.

V0, = fsolve(objective, 3, args=(0.1,))
V0

3.6764763125625461



Now, we can define the functions, and integrate them to get the same solution. I defined these pretty verbosely, just for readability.

from scipy.integrate import solve_ivp

def dPdV(V):
return -R * T / (V - b)**2 + 2 * a / V**3

def dVdP(V):
return 1 / dPdV(V)

dPdPr = Pc

def dVdPr(Pr, V):
return dVdP(V) * dPdPr

Pr_span = (0.1, 10)
Pr_eval, h = np.linspace(*Pr_span, retstep=True)

sol = solve_ivp(dVdPr, Pr_span, (V0,), dense_output=True, max_step=h)

V = sol.y[0]
P = sol.t * Pc
Z = P * V / (R * T)
plt.plot(sol.t, Z)
plt.xlabel('$P_r$')
plt.ylabel('Z')
plt.xlim([0, 10])
plt.ylim([0, 2])

(0, 2)



This also looks like Figure 7-1. It is arguably a better approach since we only need an initial condition, and after that have a reliable integration (rather than many iterative solutions from an initial guess of the solution in fsolve).

The only downside to this approach (in my opinion) is the need to derive and implement derivatives. As equations of state get more complex, this gets more tedious and complicated.

## 3 Method 3 - autograd + solve_ivp

The whole point of automatic differentiation is to get derivatives of functions that are written as programs. We explore here the possibility of using this to solve this problem. The idea is to use autograd to define the derivative $$dP/dV$$, and then solve the ODE like we did before.

from autograd import grad

def P(V):
return R * T / (V - b) - a / V**2

def dVdPr(Pr, V):
return 1 / dPdV(V) * Pc

sol = solve_ivp(dVdPr,  Pr_span, (V0,), dense_output=True, max_step=h)

V, = sol.y
P = sol.t * Pc
Z = P * V / (R * T)
plt.plot(sol.t, Z)
plt.xlabel('$P_r$')
plt.ylabel('Z')
plt.xlim([0, 10])
plt.ylim([0, 2])

(0, 2)



Not surprisingly, this answer looks the same as the previous ones. I think this solution is pretty awesome. We only had to implement the van der Waal equation, and then let autograd do its job to get the relevant derivative. We don't get a free pass on calculus here; we still have to know which derivatives are important. We also need some knowledge about how to use autograd, but with that, this problem becomes pretty easy to solve.

org-mode source

Org-mode version = 9.1.13

## A new ode integrator function in scipy

| categories: | tags: | View Comments

I learned recently about a new way to solve ODEs in scipy: scipy.integrate.solve_ivp. This new function is recommended instead of scipy.integrate.odeint for new code. This function caught my eye because it added functionality that was previously missing, and that I had written into my pycse package. That functionality is events.

To explore how to use this new function, I will recreate an old blog post where I used events to count the number of roots in a function. Spoiler alert: it may not be ready for production.

The question at hand is how many roots are there in $$f(x) = x^3 + 6x^2 - 4x - 24$$, and what are they. Now, I know there are three roots and that you can use np.roots for this, but that only works for polynomials. Here they are, so we know what we are looking for.

import numpy as np
np.roots([1, 6, -4, -24])

array([-6.,  2., -2.])



The point of this is to find a more general way to count roots in an interval. We do it by integrating the derivative of the function, and using an event function to count when the function is equal to zero. First, we define the derivative:

$$f'(x) = 3x^2 + 12x - 4$$, and the value of our original function at some value that is the beginning of the range we want to consider, say $$f(-8) = -120$$. Now, we have an ordinary differential equation that can be integrated. Our event function is simply, it is just the function value $$y$$. In the next block, I include an optional t_eval arg so we can see the solution at more points.

def fprime(x, y):
return 3 * x**2 + 12 * x - 4

def event(x, y):
return y

import numpy as np
from scipy.integrate import solve_ivp
sol = solve_ivp(fprime, (-8, 4), np.array([-120]), t_eval=np.linspace(-8, 4, 10), events=[event])
sol

 message: 'The solver successfully reached the interval end.'
nfev: 26
njev: 0
nlu: 0
sol: None
status: 0
success: True
t: array([-8.        , -6.66666667, -5.33333333, -4.        , -2.66666667,
-1.33333333,  0.        ,  1.33333333,  2.66666667,  4.        ])
t_events: [array([-6.])]
y: array([[-120.        ,  -26.96296296,   16.2962963 ,   24.        ,
10.37037037,  -10.37037037,  -24.        ,  -16.2962963 ,
26.96296296,  120.        ]])


sol.t_events

[array([-6.])]



Huh. That is not what I expected. There should be three values in sol.t_events, but there is only one. Looking at sol.y, you can see there are three sign changes, which means three zeros. The graph here confirms that.

%matplotlib inline
import matplotlib.pyplot as plt
plt.plot(sol.t, sol.y[0])

[<matplotlib.lines.Line2D at 0x151281d860>]



What appears to be happening is that the events are only called during the solver steps, which are different than the t_eval steps. It appears a workaround is to specify a max_step that can be taken by the solver to force the event functions to be evaluated more often. Adding this seems to create a new cryptic warning.

sol = solve_ivp(fprime, (-8, 4), np.array([-120]), events=[event], max_step=1.0)
sol

/Users/jkitchin/anaconda/lib/python3.6/site-packages/scipy/integrate/_ivp/rk.py:145: RuntimeWarning: divide by zero encountered in double_scalars
max(1, SAFETY * error_norm ** (-1 / (order + 1))))


 message: 'The solver successfully reached the interval end.'
nfev: 80
njev: 0
nlu: 0
sol: None
status: 0
success: True
t: array([-8.        , -7.89454203, -6.89454203, -5.89454203, -4.89454203,
-3.89454203, -2.89454203, -1.89454203, -0.89454203,  0.10545797,
1.10545797,  2.10545797,  3.10545797,  4.        ])
t_events: [array([-6., -2.,  2.])]
y: array([[-120.        , -110.49687882,  -38.94362768,    3.24237128,
22.06111806,   23.51261266,   13.59685508,   -1.68615468,
-16.33641662,  -24.35393074,  -19.73869704,    3.50928448,
51.39001383,  120.        ]])


sol.t_events

[array([-6., -2.,  2.])]



That is more like it. Here, I happen to know the answers, so we are safe setting a max_step of 1.0, but that feels awkward and unreliable. You don't want this max_step to be too small, because it probably makes for more computations. On the other hand, it can't be too large either because you might miss roots. It seems there is room for improvement on this.

It also seems odd that the solve_ivp only returns the t_events, and not also the corresponding solution values. I guess in this case, we know the solution values are zero at t_events, but, supposing you instead were looking for a maximum value by getting a derivative that was equal to zero, you might end up getting stuck solving for it some how.

Let's consider this parabola with a maximum at $$x=2$$, where $$y=2$$:

x = np.linspace(0, 4)
plt.plot(x, 2 - (x - 2)**2)

[<matplotlib.lines.Line2D at 0x1512dad9e8>]



We can find the maximum like this.

def yprime(x, y):
return -2  * (x - 2)

def maxevent(x, y):
return yprime(x, y)

sol = solve_ivp(yprime, (0, 4), np.array([-2]), events=[maxevent])
sol

/Users/jkitchin/anaconda/lib/python3.6/site-packages/scipy/integrate/_ivp/rk.py:145: RuntimeWarning: divide by zero encountered in double_scalars
max(1, SAFETY * error_norm ** (-1 / (order + 1))))


 message: 'The solver successfully reached the interval end.'
nfev: 20
njev: 0
nlu: 0
sol: None
status: 0
success: True
t: array([ 0.        ,  0.08706376,  0.95770136,  4.        ])
t_events: [array([ 2.])]
y: array([[-2.        , -1.65932506,  0.91361355, -2.        ]])



Clearly, we found the maximum at x=2, but now what? Re-solve the ODE and use t_eval with the t_events values? Use a fine t_eval array, and interpolate the solution? That doesn't seem smart. You could make the event terminal, so that it stops at the max, and then read off the last value, but this will not work if you want to count more than one maximum, for example.

maxevent.terminal = True
solve_ivp(yprime, (0, 4), (-2,), events=[maxevent])

/Users/jkitchin/anaconda/lib/python3.6/site-packages/scipy/integrate/_ivp/rk.py:145: RuntimeWarning: divide by zero encountered in double_scalars
max(1, SAFETY * error_norm ** (-1 / (order + 1))))


 message: 'A termination event occurred.'
nfev: 20
njev: 0
nlu: 0
sol: None
status: 1
success: True
t: array([ 0.        ,  0.08706376,  0.95770136,  2.        ])
t_events: [array([ 2.])]
y: array([[-2.        , -1.65932506,  0.91361355,  2.        ]])



Internet: am I missing something obvious here?

org-mode source

Org-mode version = 9.1.13

## Solving ODEs with a neural network and autograd

| categories: | tags: | View Comments

In the last post I explored using a neural network to solve a BVP. Here, I expand the idea to solving an initial value ordinary differential equation. The idea is basically the same, we just have a slightly different objective function.

$$dCa/dt = -k Ca(t)$$ where $$Ca(t=0) = 2.0$$.

Here is the code that solves this equation, along with a comparison to the analytical solution: $$Ca(t) = Ca0 \exp -kt$$.

import autograd.numpy as np

def init_random_params(scale, layer_sizes, rs=npr.RandomState(0)):
"""Build a list of (weights, biases) tuples, one for each layer."""
return [(rs.randn(insize, outsize) * scale,   # weight matrix
rs.randn(outsize) * scale)           # bias vector
for insize, outsize in zip(layer_sizes[:-1], layer_sizes[1:])]

def swish(x):
"see https://arxiv.org/pdf/1710.05941.pdf"
return x / (1.0 + np.exp(-x))

def Ca(params, inputs):
"Neural network functions"
for W, b in params:
outputs = np.dot(inputs, W) + b
inputs = swish(outputs)
return outputs

# Here is our initial guess of params:
params = init_random_params(0.1, layer_sizes=[1, 8, 1])

# Derivatives

k = 0.23
Ca0 = 2.0
t = np.linspace(0, 10).reshape((-1, 1))

# This is the function we seek to minimize
def objective(params, step):
# These should all be zero at the solution
# dCadt = -k * Ca(t)
zeq = dCadt(params, t) - (-k * Ca(params, t))
ic = Ca(params, 0) - Ca0
return np.mean(zeq**2) + ic**2

def callback(params, step, g):
if step % 1000 == 0:
print("Iteration {0:3d} objective {1}".format(step,
objective(params, step)))

step_size=0.001, num_iters=5001, callback=callback)

tfit = np.linspace(0, 20).reshape(-1, 1)
import matplotlib.pyplot as plt
plt.plot(tfit, Ca(params, tfit), label='soln')
plt.plot(tfit, Ca0 * np.exp(-k * tfit), 'r--', label='analytical soln')
plt.legend()
plt.xlabel('time')
plt.ylabel('$C_A$')
plt.xlim([0, 20])
plt.savefig('nn-ode.png')

Iteration   0 objective [[ 3.20374053]]
Iteration 1000 objective [[  3.13906829e-05]]
Iteration 2000 objective [[  1.95894699e-05]]
Iteration 3000 objective [[  1.60381564e-05]]
Iteration 4000 objective [[  1.39930673e-05]]
Iteration 5000 objective [[  1.03554970e-05]]



Huh. Those two solutions are nearly indistinguishable. Since we used a neural network, let's hype it up and say we learned the solution to a differential equation! But seriously, note that although we got an "analytical" solution, we should only rely on it in the region we trained the solution on. You can see the solution above is not that good past t=10, even perhaps going negative (which is not even physically correct). That is a reminder that the function we have for the solution is not the same as the analytical solution, it just approximates it really well over the region we solved over. Of course, you can expand that region to the region you care about, but the main point is don't rely on the solution outside where you know it is good.

This idea isn't new. There are several papers in the literature on using neural networks to solve differential equations, e.g. http://www.sciencedirect.com/science/article/pii/S0255270102002076 and https://arxiv.org/pdf/physics/9705023.pdf, and other blog posts that are similar (https://becominghuman.ai/neural-networks-for-solving-differential-equations-fa230ac5e04c, even using autograd). That means to me that there is some merit to continuing to investigate this approach to solving differential equations.

There are some interesting challenges for engineers to consider with this approach though. When is the solution accurate enough? How reliable are derivatives of the solution? What network architecture is appropriate or best? How do you know how good the solution is? Is it possible to build in solution features, e.g. asymptotes, or constraints on derivatives, or that the solution should be monotonic, etc. These would help us trust the solutions not to do weird things, and to extrapolate more reliably.