Gemini is an open-source Python package which provides scalable multi-fidelity machine learning targeting the design and discovery of functional molecules and advanced materials. (https://arxiv.org/abs/2103.03391v1)
Install Gemini from source,
git clone https://github.com/rileyhickman/gemini.git
cd matter-gemini
pip install -e .GPU use is optional. We recommend using the following
tensorflow-gpu 2.4.1
CUDA Version: 11.1
cuda-toolkit-11-1 11.1.1-1
Latest cuDNNGemini can be easily trained given 2D (# samples, # dimensions) NumPy arrays containing features (x) and targets (y) for exp and cheap datasets. Predictions using Gemini are furnished with frequentist uncertainty estimates.
from gemini import Gemini
gemini = Gemini()
gemini.train(x_exp, y_exp,
x_cheap, y_cheap)
pred_mu, pred_std = gemini.predict(x_exp_test)Gemini's predictions of expensive-to-evaluate objective functions can be used to reduce the number of expensive black-box evaluations necessary to achieve a desired target value.
The deep Bayesian optimizer Gryffin currently supports Gemini as a built-in predictive model. After installing Gemini and Gryffin,
from gryffin import Gryffin
# instantiate Gryffin
gryffin = Gryffin('config_file.json')
# optimization loop
while num_eval < budget:
samples = gryffin.recommend(observations,
proxy_observations)The Gryffin config file must include a section specifying the predictive model, i.e.
...
"predictive_model": {
"model_kind": "gemini"
},
...Alternatively, you can train Gemini in an external manner, this gives the user
greater flexibility in their expreiment. Gryffin allows for the optional passing of
a callable object to its recommend method.
from gryffin import Gryffin
from gemini import GeminiOpt as Gemini
# instantiate Gryffin
gryffin = Gryffin('config_file.json')
# instantiate Gemini
gemini = Gemini()
# optimization loop
while num_eval < budget:
if len(observations) >= 2 and len(proxy_observations) >= 2:
# construct training set with current observations
training_set = gryffin.construct_training_set(observations, proxy_observations)
# train Gemini
gemini.train(training_set['train_features'], training_set['train_targets'],
training_set['proxy_train_features'], training_set['proxy_train_targets'],
num_folds=3)
# pass callable when asking Gryffin for new samples
samples = gryffin.recommend(observations,
predictive_model=gemini)In this external trianing case, you need only provide a Gryffin config file (i.e. no predictive model entries)
- Inverse design of hybrid organic inorganic perovskites
- Inverse design of multi-component metal-oxide catalysts for the oxygen evolution reaction
- Inverse design of non-fullerene acceptor molecules for light harvesting applications
We provide methods for facile multi-fidelity data preprocessing/testing for 4 datasets reported in the literature.
dataset_perovskites(10.1038/sdata.2017.57)dataset_freesolv(10.1007/s10822-014-9747-x)dataset_photobleaching(10.1002/adma.201907801)dataset_cat_oer_1_4(10.1039/C9SC05999G)
Academic collaborations and extensions/improvements to the code are encouraged. Please reach out to Riley via email if you have questions/concerns.
- Riley J. Hickman (riley.hickman@mail.utoronto.ca)
- Florian Häse
- Matteo Aldeghi
Gemini is an open-source research software. If you use Gemini in a scientific report, please cite the following article
@misc{gemini,
title={Gemini: Dynamic Bias Correction for Autonomous Experimentation},
author={Riley J. Hickman and Florian Häse and Loïc M. Roch and Alán Aspuru-Guzik},
year={2021},
eprint={2103.03391},
archivePrefix={arXiv},
primaryClass={stat.ML}
}