ramyaguru / camd

Agent-based sequential learning software for materials discovery

Computational Autonomy for Materials Discovery (CAMD)

CAMD provides a flexible software framework for sequential / Bayesian optimization type campaigns for materials discovery. Its key features include:

• Agents: Decision making entities which select experiments to run from pre-determined candidate sets. Agents can combine machine learning with physical or chemical constructs, logic, heuristics, exploration-exploitation strategies and so on. CAMD comes with several generic and structure-discovery focused agents, which can be used by the users as templates to derive new ones.
• Experiments: Entities responsible for carrying out the experiments requested by Agents and reporting back the results.
• Analyzers: Post-processing procedures which frame experimental results in the context of candidate or seed datasets.
• Campaigns: Loop construct which executes the sequence of hypothesize-experiment-analyze by the Agent, Experiment, and Analyzer, respectively, and facilitates the communication between these entities.
• Simulations: Agent performance can be simulated using after-the-fact sampling of known existing data. This allows systematic design and tuning of agents before their deployment using actual Experiments.

A more in-depth description of the scientific framework can be found in this recent open-access article, which demonstrates an end-to-end CAMD-based framework for autonomous inorganic materials discovery using cloud-based density functional theory calculations.

To get started with CAMD, we recommend exploring the tutorial from the recent TRI-AMDD hackathon event available here.

Installation

CAMD can be installed using pip as pip install camd. If issues are encountered, we recommend following the installation procedures below:

Note that, since qmpy is currently only python 2.7 compatible, CAMD python 3 compatibility depends on a custom fork of qmpy here, which is installed using the setup.py procedure.

We recommend using Anaconda python, and creating a fresh conda environment for the install (e. g. conda create -n MY_ENV_NAME).

Linux

Install numpy via pip first, since the build depends on this and numpy has some difficulty recognizing its own install. Then install requirements and use setup.py.

pip install numpy
pip install -r requirements.txt
python setup.py develop

Mac OSX

First dependencies via homebrew. Thanks to the contributors to this stack exchange thread.

brew install gcc

Install numpy via pip first, since the build depends on this and numpy has some difficulty recognizing its own install. Then install requirements and use setup.py.

pip install numpy
pip install -r requirements.txt
python setup.py develop

Data download

Datasets for featurized OQMD entries for after-the-fact testing can be downloaded at data.matr.io/3. These are done automatically in the code and stored in the camd/_cache directory.

Citation

If you use CAMD, we kindly ask you to cite the following publication:

• Montoya, J. H., Winther, K. T., Flores, R. A., Bligaard, T., Hummelshøj, J. S., & Aykol, M. "Autonomous intelligent agents for accelerated materials discovery" Chemical Science 11 (2020) 8517–8532 doi:10.1039/D0SC01101K, open-access.