Theory

ESPEI has two types of fitting – parameter generation and MCMC optimization. The parameter generation step uses experimental and DFT data of the Gibbs free energy derivatives (\(C_P, H, S\)) for each phase and for the mixing energies within sublattices for each phase to generate and fit parameters of given CALPHAD models. The MCMC optimization step uses a Bayesian optimization procedure to fit parameters in a Database to experimental phase equilibria.

Parameter generation

A simple model with few parameters is better than a complex model that describes the same data marginally better. Parameter generation in ESPEI aims to achieve a balance of a simple parameterization and goodness of fit in the Redlich-Kister polynomial used in CALPHAD assessments. To achieve this, parameters are selected using the corrected Akaike information criterion (AICc) to choose an optimal set of parameters from canditate parameterizations.

The general Redlich Kister polynomial has the form \(G = a + bT + cT\ln T + \sum_n d_n T^n\). Different parameterizations, e.g. only considering \(a\), considering \(a\) and \(b\), \(a\) and \(c\), etc. are fit to all of the input formation or mixing data (depending on the parameter being selected) by a least squares pseudo-inverse optimization.

Each parameterization is compared in the AICc and the most suitable optimization balances the goodness of fit and the number of parameters. The key aspect of this is that ESPEI will avoid overfitting your data and will not add parameters you do not have data for.

This is important for phases that would have a temperature dependent contribution to the Gibbs energy, but the input data only gives 0K formation energies. ESPEI cannot add temperature dependence to the parameterized model. Because of this, an abundance of single-phase data is critical to provide enough degrees of freedom in later optimization.

MCMC optimization

Details of Markov Chain Monte Carlo as an algorithm are better covered elsewhere. A good example is MacKay’s (free) book: Information Theory, Inference, and Learning Algorithms.

Using MCMC for optimizing CALPHAD models might appear to have several drawbacks. The parameters in the models are correlated and due to the nature of single phase first-principles data the shape and size of the posterior distribution for each parameter is not known before fitting. Traditional Metropolis-Hastings MCMC algorithms require the a prior to be defined for each parameter, which is a challenge for parameters in CALPHAD models which vary over more than 6 orders of magnitude.

ESPEI solves these potential problems by using an Ensemble sampler, as introduced by Goodman and Weare [1], rather than the Metropolis-Hastings algorithm. Ensemble samplers have the property of affine invariance, which uses multiple (\(\geq 2 N\) for \(N\) parameters) parallel chains to scale new proposal parameters by linear transforms. These chains, together an ensemble, define a proposal distribution to sample parameters from that is scaled to the magnitude and sensitivity of each parameter. Thus, Ensemble samplers directly address the challenges we expect to encounter with traditional MCMC.

ESPEI uses an Ensemble sampler algorithm by using the emcee package that implements parallelizable ensemble samplers. To use emcee, ESPEI defines the initial ensemble of chains and a function that returns the error as a log-probability. ESPEI defines the error as the mean square error between experimental phase equilibria and the equilibria calculated by the CALPHAD database.

Here, again, it is critical to point out the importance of abundant phase equilibria data. Traditional CALPHAD modeling has involved the modeler participating in tight feedback loops between updates to parameters and the resulting phase diagram. ESPEI departs from this by optimizing just a single scalar error function based on phase equilibria. The implication of this is that if there are phase equilibria that are observed to exist, but they are not in the datasets that are considered by ESPEI, those equilibria cannot be optimized against and may deviate from ‘known’ equilibria. A possible approach to address this in ESPEI is to estimate the points for the equilibria.

References

[1]Goodman, J. & Weare, J. Ensemble samplers with affine invariance. Commun. Appl. Math. Comput. Sci. 5, 65–80 (2010). doi:10.2140/camcos.2010.5.65.