Table 1.

Benchmarking our deep learning model–ElemNet–against conventional machine learning approaches.

Model	Input Type	MAE (eV/atom)	Training time (hour)	Prediction time (sec)
RandomForest	Physical Attributes	0.071 ± 0.0006	1.5	14.80
RandomForest	Elemental Compositions	0.157 ± 0.0012	1.5	2.87
ElemNet	Elemental Compositions	0.050 ± 0.0007	7 (GPU)	9.28 (CPU) & 0.08 (GPU)

We trained several conventional ML models such as Linear Regression, SGDRegression, ElasticNet, AdaBoost, Ridge, RBFSVM, DecisionTree, ExtraTrees, Bagging and Random Forest. Out of them, Random Forest performed the best with and without using physical attributes. Here, we show the results from our deep learning model and the best conventional ML model- Random Forest, in our study for both types of model inputs (without and without physical attributes), along with the type of input used, mean absolute error (MAE) on the test set, training time on the training set, and prediction time on the entire test set (25,662 entries). All the models are trained and tested using a ten-fold cross validation. All timings are on a single (logical) CPU core of an NVIDIA DIGITS DevBox with a Core i7-5930K 6 Core 3.5 GHz desktop processor with 64GB DDR4 RAM and 4 TITAN X GPUs with 12GB of memory per GPU, except the deep learning models.