Surrogate-enabled Bayesian sampling of force field parameters
Posted on 3 Sep 2021 by Owen Madin
In the Open Force Field Initiative, we are always looking for novel ways to advance force field parameterization. Working on the parameterization of Lennard-Jones models, one strategy I am exploring is using Bayesian inference to explore parameter landscapes with respect to physical properties. This is a lot to parse in one sentence, so let’s break it down!
Lennard-Jones parameterization relies on physical properties
Since the earliest days of force field science, LJ parameters have been fitted against physical property datasets ( typically liquid densities and heats of vaporization). This is an important step of parameterization because it fits the critical LJ parameters and introduces macroscopic constraints that complement microscopic QM constraints.
The disadvantage of using physical property measurements in parameterization is that they are generally slow to evaluate, because computing a physical property with a force field requires you to run an equilibrium simulation with that force field. This makes it difficult to efficiently optimize LJ parameters or otherwise explore parameter space due to computational constraints.
Bayesian posterior sampling is a promising method of exploring parameter space
Bayesian inference is a paradigm for evaluating parameter sets that naturally incorporates prior information and a likelihood derived from experimental data into a posterior distribution, which can be used as a metric for fitness. Bayesian inference is of interest to us because of a) the way that it parsimoniously incorporates previous model knowledge, and b) because understanding the whole parameter distribution can find parameter sets that escape local minima.
The catch, however, is that evaluating a Bayesian posterior distribution generally requires Monte Carlo simulation, and is even more computationally expensive then most optimization schemes. So if we want to do Bayesian exploration of a parameter probability distribution, we need a way to speed things up significantly. For us, this means we need a much faster way of taking LJ parameter inputs (sigma/rmin_half, epsilon) and getting physical property outputs (densities, enthalpies) very quickly.
We can accelerate sampling with surrogate modeling
Fortunately, the physical properties we are interested in are relatively smooth functions of the LJ parameters in our regions of interest, and can be approximated with simpler surrogate models with limited losses of fidelity. For our surrogate models, we can use Gaussian Processes, which are simple surrogates that excel at making predictions (with uncertainty) from a limited number of observations. The technique originated in geostatistics, where it is desirable to understand mineral content over a wide area of space, but only feasible to take a limited number of samples. The principle is the same here!
Putting it together
In the rest of this post, we’ll demonstrate:
- Using the OpenFF evaluator, automatically simulating and collating physical property measurements from simulation at a variety of parameter sets.
- Using GPytorch, building a surrogate model based off the outputs and uncertainties of these simulated measurements.
- Using the Pyro probabilistic programming language to model a Bayesian posterior distribution, and then sampling from this distribution using the NUTS MCMC sampler to get the parameter landscape.
For the purpose of this demonstration, we’ll focus on a simple example (easy to visualize!): building a surrogate model
for the density of a 50/50 mixture of pentane and hexane (at STP), as a function of the epsilon and rmin_half of the
tetravalent carbon parameter [#6X4:1]
.
Simulating the density with respect to parameters
To start, we focus on the line from the OpenFF 1.3.0 .offxml
that contains the [#6X4:1]
(tetravalent carbon) LJ
parameters:
<Atom smirks="[#6X4:1]" epsilon="0.1094 * mole**-1 * kilocalorie" id="n16" rmin_half="1.908 * angstrom"/>
Using the OpenFF toolkit, it’s easy to access and modify these parameters. In order to come up with some efficiently spaced parameter sets, we use Latin hypercube sampling (from the SMT package) to get samples from the range of 90-110% of the original parameters.
import numpy as np
from smt.sampling_methods import LHS
xlimits = np.asarray([[0.9, 1.1], [0.9, 1.1]])
# Set the limits to sample within 90-110% of the original parameters (in 2D)
hypercube = LHS(xlimits=xlimits)
# Create LHS sampling object
num_samples = 40
samples = hypercube(num_samples)
# Create 40 samples from the hypercube
Here are the first 5 resulting hypercube samples:
array([[1.0725, 1.0025],
[1.0275, 1.0925],
[1.0125, 0.9425],
[0.9225, 0.9475],
[0.9825, 0.9175],
Taking the first Latin hypercube sample, we can easily use it to modify the force field:
from openff.toolkit.typing.engines.smirnoff.forcefield import ForceField
forcefield = ForceField('openff-1-3-0.offxml', allow_cosmetic_attributes=True)
lj_params = forcefield.get_parameter_handler('vdW', allow_cosmetic_attributes=True)
# Get the forcefield from file and its LJ parameters
lhs_sample = [1.0725, 1.0025]
smirks_to_change = '[#6X4:1]'
lj_params[smirks_to_change].epsilon *= lhs_sample[0]
lj_params[smirks_to_change].rmin_half *= lhs_sample[1]
# Multiply the parameters of interest by the values from the LHS sample
forcefield.to_file('force-field.offxml')
# Save the result to file
print(lj_params[smirks_to_change])
# Show the changed parameters
<vdWType with smirks: [#6X4:1] epsilon: 0.12583803375 kcal/mol id: n16 rmin_half: 1.9127699999999999 A >
With this method, we can easily generate many perturbed force fields to sample from. The next thing we need to do is simulate our property of interest with all these forcefields using the OpenFF Evaluator framework!
We are going to spin up an Evaluator server on our compute resources and then pass an estimation request that corresponds to each of the parameter sets to that server. First we define a function that passes a request to the server:
from openff.evaluator.properties import Density, EnthalpyOfMixing
from openff.evaluator.client import RequestOptions
from openff.evaluator.client import EvaluatorClient
def estimate_forcefield_properties(property_dataset, forcefield):
data_set = property_dataset
force_field_source = forcefield
density_schema = Density.default_simulation_schema(n_molecules=1000)
h_mix_schema = EnthalpyOfMixing.default_simulation_schema(n_molecules=1000)
# Create an options object which defines how the data set should be estimated.
estimation_options = RequestOptions()
# Specify that we only wish to use molecular simulation to estimate the data set.
estimation_options.calculation_layers = ["SimulationLayer"]
# Add our custom schemas, specifying that the should be used by the 'SimulationLayer'
estimation_options.add_schema("SimulationLayer", "Density", density_schema)
estimation_options.add_schema("SimulationLayer", "EnthalpyOfMixing", h_mix_schema)
evaluator_client = EvaluatorClient()
request, _ = evaluator_client.request_estimate(
property_set=data_set,
force_field_source=force_field_source,
options=estimation_options,
)
return request
Then we can create a function to spin up the evaluator server that we pass the requests to:
from openff.evaluator.backends import QueueWorkerResources
from openff.evaluator.backends.dask import DaskLSFBackend
from openff.evaluator.server import EvaluatorServer
from openff.evaluator.datasets import PhysicalPropertyDataSet
from openff.evaluator.forcefield import SmirnoffForceFieldSource
# Set up a backend to run the calculations on with the requested resources.
cpus_per_worker = 1
gpus_per_worker = 5
worker_resources = QueueWorkerResources(
number_of_threads=cpus_per_worker,
number_of_gpus=gpus_per_worker,
preferred_gpu_toolkit=QueueWorkerResources.GPUToolkit.CUDA,
)
lsf_backend = DaskLSFBackend(minimum_number_of_workers=1,
maximum_number_of_workers=50,
resources_per_worker=worker_resources)
files_directory = 'modified-force-fields'
working_directory = 'working-directory'
with lsf_backend:
with EvaluatorServer(calculation_backend=lsf_backend,
working_directory=working_directory,
port=8000,
enable_data_caching=False,
delete_working_files=True):
requests = []
forcefields = []
for subdirectory in os.listdir(files_directory):
forcefield = SmirnoffForceFieldSource.from_path(
os.path.join(files_directory, subdirectory, 'force-field.offxml'))
property_dataset = PhysicalPropertyDataSet.from_json(
os.path.join(files_directory, subdirectory, 'test-set-collection.json'))
requests.append(estimate_forcefield_properties(property_dataset, forcefield))
forcefields.append(forcefield)
results = [
request.results(synchronous=True, polling_interval=30)[0]
for request in requests
]
In this case, the test-set-collection.json
object contains the physical property data set that we are interested in
simulating. In our case, this is a single density data point, sourced from NIST ThermoML:
Temperature (K) | Pressure (kPa) | Density Value (g / ml) | Density Uncertainty (g / ml) | Component 1 | Mole Fraction 1 | Component 2 | Mole Fraction 2 |
---|---|---|---|---|---|---|---|
298.15 | 101 | 0.63966 | 0.000355 | CCCCC | 0.5 | CCCCCC | 0.5 |
We spin up the server, pass requests to simulate this data point with each modified force field, and then collect the results of the simulations. After some data collation and formatting (not shown here), we’ll be ready to build a surrogate model! Before doing so, we can take a peek at some of the results:
Parameters:
[#6X4:1]_epsilon [#6X4:1]_rmin_half
0 0.119520 1.97001
1 0.111315 1.78875
2 0.118973 2.09403
3 0.101469 1.93185
4 0.106391 1.76967
Density values:
CCCCCC{solv}{x=0.500000}|CCCCC{solv}{x=0.500000}_density
0 0.666314
1 0.570434
2 0.670585
3 0.620133
4 0.536805
Building the surrogate models
With our data ready, we can build a GP surrogate model using the GPytorch package. To formulate our surrogate model, we use the squared exponential covariance kernel, and the simple constant mean function. For our likelihood function, we choose a Gaussian likelihood. This implementation is taken from this example from the GPytorch documentation.
import gpytorch
class ExactGPModel(gpytorch.models.ExactGP):
def __init__(self, train_x, train_y, likelihood):
super(ExactGPModel, self).__init__(train_x, train_y, likelihood)
# Choose constant mean function for GP
self.mean_module = gpytorch.means.ConstantMean()
# Choose RBF/Squared Exponential covariance kernel for GP
self.covar_module = gpytorch.kernels.ScaleKernel(gpytorch.kernels.RBFKernel())
def forward(self, x):
mean_x = self.mean_module(x)
covar_x = self.covar_module(x)
return gpytorch.distributions.MultivariateNormal(mean_x, covar_x)
likelihood = gpytorch.likelihoods.GaussianLikelihood()
# Choose simple Gaussian likelihood model for GP
model = ExactGPModel(parameters, densities, likelihood)
In this case, the parameters
and densities
that we have passed to the surrogate model have been converted to torch
tensors to work with GPytorch
. Now that we’ve initialized our surrogate model, we need to train its hyperparameters
and get it ready to evaluate:
# Set our surrogate model into training mode
model.train()
likelihood.train()
# Train for 1000 iterations
training_iter = 1000
# Use the adam optimizer
optimizer = torch.optim.Adam(model.parameters(), lr=0.1) # Includes GaussianLikelihood parameters
# Set "Loss" for GPs - the marginal log likelihood
mll = gpytorch.mlls.ExactMarginalLogLikelihood(likelihood, model)
for i in range(training_iter):
# Zero gradients from previous iteration
optimizer.zero_grad()
# Output from model
output = model(X)
# Calc loss and backprop gradients
loss = -mll(output, Y)
loss.backward()
optimizer.step()
model.eval()
likelihood.eval()
Now, our model is ready to go! To get a sense for the quality of our surrogates, I’ve plotted the surrogate values over the range 75-125% of the original parameter values:
As we can see, the model we’ve built is a relatively smooth surface of the parameters of interest. To get an idea of the quality of the surrogate models, we can plot the surrogate model, with the points used for training superimposed over the uncertainty surface:
We see that in the region we sampled, the uncertainty of the model is quite low, and this is a high-quality surrogate.
We also note that the quality of the model degrades as we move away from the sampled regions, particularly
as rmin_half
is varied. This surrogate should be sufficient to explore the parameter landscape in the region of
interest (90-110% of the original parameter values).
Bayesian sampling of the surrogate model
To build our Bayesian posterior model based on the surrogate, we utilize the Pyro probabilistic programming language to represent our Bayesian model, and subsequently do MCMC sampling over the parameter space.
Our posterior distribution is composed of a prior and likelihood function:
Prior: Gaussian distribution for each parameter centered at the initial parameter values from OpenFF 1.3.0, with variance set to 1/10th of the mean:
Likelihood: Gaussian distribution for each simulated data point (only one in this case), centered at the experimental value, and with variance equal to the sum of the surrogate and experimental variances.
In this equation, the vector of parameter is represented by theta, densities are represented by rho, and N(x,mu,sigma^2) represented a normal PDF with mean mu and variance sigma^2.
We can represent this posterior distribution in Pyro:
import pyro
import torch
def pyro_model():
# Import initial parameters from OpenFF 1.3.0
initial_parameters = torch.tensor([[0.1094, 1.9080]])
experimental_values = torch.tensor([[0.63966]])
uncertainty_values = torch.tensor([[0.000355]])
# Place normal priors centered at the initial parameters on the LJ parameters:
parameters = pyro.sample(
"parameters",
pyro.distributions.Normal(
# Use a normal distribution with a relatively low uncertainty
# to avoid sampling in regions where the surrogate is poor
initial_parameters,
initial_parameters * 0.1,
),
)
# return evaluated surrogate values and standard deviations
evaluated_surrogate = likelihood(model(parameters))
predictions, predicted_uncertainties = evaluated_surrogate.mean, evaluated_surrogate.stddev
uncertainty = pyro.deterministic(
"uncertainty",
torch.sqrt(uncertainty_values ** 2 + predicted_uncertainties ** 2))
# Sample from posterior distribution
return pyro.sample(
"predicted_residuals",
pyro.distributions.Normal(loc=predictions, scale=uncertainty),
obs=experimental_values,
)
Now that we have a defined surrogate model, we can easily do MCMC sampling with NUTS (No U-Turn Sampler) in pyro:
from pyro.infer import MCMC, NUTS
import numpy as np
# Initialize the NUTS kernel
nuts_kernel = NUTS(pyro_model)
# Run 1 chain of the MCMC process for 10000 samples
num_samples = 10000
mcmc = MCMC(nuts_kernel,
num_samples=num_samples,
warmup_steps=int(np.floor(num_samples / 5)),
num_chains=1)
mcmc.run()
mcmc.summary()
The simulation runs and we get some output statistics:
mean std median 5.0% 95.0% n_eff r_hat
parameters[0,0] 0.11 0.01 0.10 0.10 0.11 832.57 1.00
parameters[0,1] 2.07 0.09 2.06 1.93 2.21 616.16 1.00
Number of divergences: 159
Plotting the traces, we can visualize the parameter posterior distributions:
This gives us insight into the best values of these parameters for this data set (a single point in this case), and gives us an idea of the parameter landscape.
Summary
In this blog post, we:
- Modified the LJ
[#6X4:1]
rmin_half
andepsilon
parameters using the OpenFF Toolkit, in order to get samples to build a surrogate model from. - Used the OpenFF evaluator framework to evaluate a density measurement for a 50/50 pentane/hexane mixture at each of the modified force fields.
- Built a surrogate model from the results of the simulations, using the GPytorch surrogate modeling library.
- Constructed a Bayesian posterior distribution over the parameters of interest, using the pyro library
- Used the NUTS MCMC algorithm to sample this posterior distribution.
This fast posterior sampling using surrogate models paves the way for us to implement Bayesian inference over force fields, which should help us discover better values of LJ parameters. Of course, the parameter sets and data sets we will use will be much larger, coming with scale-up challenges. However, this implementation gives us a platform to expand the scope of our sampling with the excellent tools developed in the OpenFF and elsewhere!