Materials Knowledge Systems in Python



The Materials Knowledge Systems (MKS) is a novel data science approach for solving multiscale materials science problems. It uses techniques from physics, machine learning, regression analysis, signal processing, and spatial statistics to create structure-property-processing relationships. The MKS carries the potential to bridge multiple length scales using localization and homogenization linkages, and provides a data driven framework for solving inverse material design problems.

See these references for further reading:

  • Computationally-Efficient Fully-Coupled Multi-Scale Modeling of Materials Phenomena Using Calibrated Localization Linkages, S. R. Kalidindi; ISRN Materials Science, vol. 2012, Article ID 305692, 2012, doi:10.5402/2012/305692.
  • Formulation and Calibration of Higher-Order Elastic Localization Relationships Using the MKS Approach, Tony Fast and S. R. Kalidindi; Acta Materialia, vol. 59 (11), pp. 4595-4605, 2011, doi:10.1016/j.actamat.2011.04.005
  • Developing higher-order materials knowledge systems, T. N. Fast; Thesis (PhD, Materials engineering)–Drexel University, 2011, doi:1860/4057.


The Materials Knowledge Materials in Python (PyMKS) framework is an object oriented set of tools and examples written in Python that provide high level access to the MKS framework for rapid creation and analysis of structure-property-processing relationships. A short intoduction of how to use PyMKS is outlined below and example cases can be found in the examples section. Both code and example contributions are welcome.

Mailing List

Please feel free to ask open ended questions about PyMKS on the list.

Meet PyMKS

In this short introduction, we will demonstrate the functionality of PyMKS to compute 2-point statistics in order to objectively quantify microstructures, predict effective properties using homogenization and predict local properties using localization. If you would like more technical details amount any of these methods please see the theory section.

%matplotlib inline
%load_ext autoreload
%autoreload 2

import numpy as np
import matplotlib.pyplot as plt

Quantify Microstructures using 2-Point Statistics

Lets make two dual phase microstructures with different morphologies.

from pymks.datasets import make_microstructure

X_1 = make_microstructure(n_samples=1, grain_size=(25, 25))
X_2 = make_microstructure(n_samples=1, grain_size=(15, 95))

X = np.concatenate((X_1, X_2))

Throughout PyMKS X is used to represent microstructures. Now that we have made the two microstructures, lets take a look at them.

from import draw_microstructures


We can compute the 2-point statistics for these two periodic microstructures using the correlate function from pymks.stats. This function computes all of the autocorrelations and cross-correlation(s) for a microstructure. Before we compute the 2-point statistics, we will discretize them using the PrimitiveBasis function.

from pymks import PrimitiveBasis
from pymks.stats import correlate

prim_basis = PrimitiveBasis(n_states=2, domain=[0, 1])
X_ = prim_basis.discretize(X)
X_corr = correlate(X_, periodic_axes=[0, 1])

Let’s take a look at the two autocorrelations and the cross-correlation for these two microstructures.

from import draw_correlations

print X_corr[0].shape

(101, 101, 3)

2-Point statistics provide an object way to compare microstructures, and have been shown as an effective input to machine learning methods.

Predict Homogenized Properties

In this section of the intro, we are going to predict the effective stiffness for two phase microstructures using the MKSHomogenizationModel, but we could have chosen any other effective material property.

First we need to make some microstructures and their effective stress values to fit our model. Let’s create 200 random instances 3 different types of microstructures, totaling to 600 microstructures.

from pymks.datasets import make_elastic_stress_random

grain_size = [(47, 6), (4, 49), (14, 14)]
n_samples = [200, 200, 200]

X_train, y_train = make_elastic_stress_random(n_samples=n_samples, size=(51, 51),
                                              grain_size=grain_size, seed=0)

Once again, X_train is our microstructures. Throughout PyMKS y is used as either the prpoerty or the field we would like to predict. In this case y_train is the effective stress values for X_train. Let’s look at one of each of the three different types of microstructures.


The MKSHomogenizationModel uses 2-point statistics, so we need provide a discretization method for the microstructures by providing a basis function. We will also specify which correlations we want.

from pymks import MKSHomogenizationModel

prim_basis = PrimitiveBasis(n_states=2, domain=[0, 1])
homogenize_model = MKSHomogenizationModel(basis=prim_basis,
                                          correlations=[(0, 0), (1, 1), (0, 1)])

Let’s fit our model with the data we created., y_train, periodic_axes=[0, 1])

Now let’s make some new data to see how good our model is.

n_samples = [10, 10, 10]
X_test, y_test = make_elastic_stress_random(n_samples=n_samples, size=(51, 51),
                                            grain_size=grain_size, seed=100)

We will try and predict the effective stress of our X_test microstructures.

y_pred = homogenize_model.predict(X_test, periodic_axes=[0, 1])

The MKSHomogenizationModel generates low dimensional representations of microstructures and regression methods to predict effective properties. Take a look at the low dimensional representations.

from import draw_components

draw_components(homogenize_model.fit_data, homogenize_model.predict_data,
                label_1='Training Data', label_2='Testing Data')

Now let’s look at a goodness of fit plot for our MKSHomogenizationModel.

from import draw_goodness_of_fit

fit_data = np.array([y_train,
                     homogenize_model.predict(X_train, periodic_axes=[0, 1])])
pred_data = np.array([y_test, y_pred])

draw_goodness_of_fit(fit_data, pred_data, ['Training Data', 'Testing Data'])

Looks good.

The MKSHomogenizationModel can be used to predict effective properties and processing-structure evolutions.

Predict Local Properties

In this section of the intro, we are going to predict the local strain field in a microstructure using MKSLocalizationModel, but we could have predicted another local property.

First we need some data, so let’s make some.

from pymks.datasets import make_elastic_FE_strain_delta

X_delta, y_delta = make_elastic_FE_strain_delta()

Once again, X_delta is our microstructures and y_delta is our local strain fields. We need to discretize the microstructure again so we will also use the same basis function.

from pymks import MKSLocalizationModel

prim_basis = PrimitiveBasis(n_states=2)
localize_model = MKSLocalizationModel(basis=prim_basis)

Let’s use the data to fit our MKSLocalizationModel., y_delta)

Now that we have fit our model, we will create a random microstructure and compute its local strain field using finite element analysis. We will then try and reproduce the same strain field with our model.

from pymks.datasets import make_elastic_FE_strain_random

X_test, y_test = make_elastic_FE_strain_random()

Let’s look at the microstructure and its local strain field.

from import draw_microstructure_strain

draw_microstructure_strain(X_test[0], y_test[0])

Now let’s pass that same microstructure to our MKSLocalizationModel and compare the predicted and computed local strain field.

from import draw_strains_compare

y_pred = localize_model.predict(X_test)
draw_strains_compare(y_test[0], y_pred[0])

Not bad.

The MKSLocalizationModel can be used to predict local properties and local processing-structure evolutions.