/emmet

Be a master builder of databases of material properties. Avoid the Kragle.

Primary LanguagePythonOtherNOASSERTION

Emmet

The purpose of Emmet is to 'build' collections of materials properties from the output of computational materials calculations. Currently, the effective purpose of Emmet is to take the output of VASP electronic structure calculations to build MongoDB collections that back the Materials Project website and its apps.

Emmet uses Maggma, our more general aggregation framework which abstracts away the behind-the-scenes machinery: Maggma provides our Builder class and a general interface to Stores, which can be MongoDB collections or plain JSON files.

The Builder takes source Store(s), processes items in that store, and then builds results to target Store(s).

To ease debugging, in Emmet data flows in one direction only: this means that each Store is only built by a specific builder, and will not then be modified by successive builders.

Builders are designed to be run periodically and automatically: as such, Stores have a 'last updated' filter (lu_filter) so that we only attempt to process new entries in the Store.

Emmet is currently primarily an internal Materials Project tool so if you're reading this you might have just joined the group, in which case: welcome! :-)

Table of Contents

Installation

Emmet is on PyPI, so pip install emmet should work. However, it is currently in very active development, so an editable install is recommended for now:

pip install -e git+git@github.com:materialsproject/emmet.git#egg=emmet --src .

Running a Builder

Here is a sample script for running the MaterialsBuilder. Replace database information as appropriate (this assumes a test database running on localhost with a pre-populated tasks collection, with mat.json in the working directory).

#!/usr/bin/env python

from maggma.runner import Runner
from maggma.stores import MongoStore, JSONStore
from emmet.vasp.builders.materials import MaterialsBuilder
from emmet.vasp.builders.thermo import ThermoBuilder

tasks_store = MongoStore(database="test",
                         collection_name="materials",
                         host="localhost",
                         port=27017,
                         lu_field="last_updated")
materials_settings_store = JSONStore("mat.json")
materials_store = MongoStore(database="test",
                             collection_name="tasks",
                             host="localhost",
                             port=27017)

materials_builder = MaterialsBuilder(tasks_store,
                                     materials_settings_store,
                                     materials_store,
                                     lu_field="last_updated")

runner = Runner([materials_builder])

runner.run()

Take care to set the lu_field correctly: this is the key that the builder looks for to see when the document was last updated, and thus which new documents to build from. This field does not exist by default in MongoDB.

To run more than one builder, add:

thermo_store = MongoStore(database="test",
                          collection="thermo",
                          host="localhost",
                          port=27017)
                          
thermo_builder = ThermoBuilder(materials_store,
                               thermo_store)

and change runner = Runner([materials_builder]) to runner = Runner([materials_builder, thermo_builder]).

The list of builders can be provided in any order: their dependencies will be resolved intelligently and the Runner will run the builders in the correct order and in parallel if supported by the system.

Writing a New Builder

Sub-class the Builder base class and implement the following methods:

  • get_items() – get your items to process, e.g. as a result of a running a query on your source(s)
  • process_item() – for each of your items, do something, e.g. calculate a diffraction pattern
  • update_targets() – update your target(s) with your processed data
  • finalize() – optional, perform any final clean up (close database connections etc., the base class can handle this)

The DiffractionBuilder is a nice simple builder to copy from to get started.

VASP Builders

The VASP builders all operate on a tasks Store which is parsed from any VASP calculation folder by Atomate's VaspDrone. Once the tasks Store has been created, Emmet's builders take over.

Overview Flowchart: Vasp Output Directory leads to Tasks Store (via VaspDrone, atomate.vasp.drones), Tasks Store with Materials Setting Store and StructureNLs Store leads to Materials Store (via MaterialsBuilder, emmet.vasp.builders.materials), Materials Store leads to Thermo Store (via ThermoBuilder, emmet.vasp.builders.thermo), Materials Store leads to Elastic Store (via ElasticBuidler, emmet.vasp.builders.elastic), Materials Store leads to Diffraction Store (via DiffractionBuilder, emmet.vasp.builders.diffraction), Materials Store leads to Dielectric Store (via DielectricStore, emmet.vasp.builders.dielectric)

MaterialsBuilder

Source(s) tasks (typically tasks collection), material_settings (typically mat.json), snls (optional)

Target(s) materials (typically materials collection)

What MaterialsBuilder does:
  1. Filters to only include tasks that completed successfuly.

  2. Groups tasks into those for the same structure.

    Structure matching first only selects materials that have the same chemical formula, and then uses pymatgen's StructureMatcher to perform symmetry analysis.

  3. For each property, ranks tasks for a given structure according to those that are expected to predict the property more accurately (for example, a band gap from a band structure calculation is ranked higher than a band gap from a generic calculation). This value is then picked as the canonical value for that property.

    The task_type is already determined and comes from the tasks store, and the rankings are specified in mat.json. No attempt is made to rank which task of the same task_type is best; in this case it is assumed that the most recent calculation takes precendence.

  4. (Optional) The Structure Notation Language (or 'SNLs') provide a way to bundle a structure and its metadata (such as bibtex references for where the structure came from) within the Materials Project. This will lookup if there are existing SNL(s) for the structure, and assign an SNL accordingly.

ThermoBuilder

Source(s) materials

Target(s) thermo

What ThermoBuilder does:
  1. Groups materials into those in the same chemical system (that is, materials whose crystal structure contain the same elements).

  2. Filters out materials that can not be directly compared to each other, e.g. they've been calculated by different methods such that their total energies are on different scales.

    By default, this is done by using MaterialsProjectCompatibility('Advanced') in pymatgen, which intelligently mixes GGA and GGA+U calculations depending on the elements present, and performs corrections to the total energy as appropriate.

  3. Uses pymatgen's phasediagram package to calculate the energy above hull for each material and, if the material is unstable, its decomposition pathway.

ElasticBuilder

Source(s) materials

Target(s) elastic

What ElasticBuilder does:
  1. Selects an initial structure from materials
  2. Finds deformed instances of this initial structure from materials, and calculates the deformation matrix
  3. If 6 independent deformations are found, calculates the elastic tensor using pymatgen's ElasticTensor

Diffraction Builder

Source(s) materials, xrd_settings (typically xrd.json)

Target(s) diffraction

What DiffractionBuilder does:
  1. For each structure, calculates its ideal X-ray diffraction pattern for a variety of X-ray targets (specified in xrd.json)

Topology Builder

Source(s) tasks, materials

Target(s) toplogy, bader

What TopologyBuilder does:
  1. For each structure in materials, calculates bonding from the material's crystal structure using a variety of methods (pymatgen's local_env and critic2's sum of atomic charge densities).

  2. It then finds the task corresponding to a static calculation.

  3. If AECCAR0, AECCAR2, CHGCAR are present, performs attempts to find bonding information using critic2 and also performs a bader analysis that is stored separately.