06 22 Meeting Summary - mp-interns/eratosthenes GitHub Wiki

Overview

Apply GAs (with a given objective function), and subsequently other algorithms, to the ~2,000 compound search space of compounds whose elastic constants have already been computed in VASP. The performance of an algorithm can be benchmarked with the already calculated values. If successful, the guessing algorithm(s) can be projected to the search space of the several thousand compounds whose elastic constants are yet to be calculated, in order to wisely select which compounds may be most promising to simulate.

Patrick's Modules

1) Define the desired behavior of a compound, and the constraint metrics that will be used to express it.

In the case of the water-splitting paper, the analogue to this module would be the particular band gap and band edge values characteristic of a desirable water-splitting compound. For the elastic constants, these constraints could pertain to something like a bulk modulus range of interest, or an element of the stiffness tensor.

Ideas: Poisson ratio, bulk/shear modulus,

2) Parameterize the search space candidates

This involves the 'labeling' of each candidate compound, and should reflect the structure in some way such that offspring generation/mutations are both physical and easily expressed.

Ideas: Space group, volume per atom(s),

3) Build a functioning algorithm

It seems we'll be starting with a genetic algorithm (likely inspired by Anubhav's modified pyevolve GA), with the prospect of other interpolation/commercial techniques later on.

Ideas:

Collaborative workflow

see separate page.

Upcoming

Meeting on Friday 06/26