Skip to main content
. 2021 Feb 19;11:4244. doi: 10.1038/s41598-021-83193-1

Figure 3.

Figure 3

Prediction error analysis for prediction formation enthalpy in JARVIS dataset using different models. We use the 145 physical attributes derived from material composition as the model inputs. We benchmark against plain network and several traditional ML models such as Linear Regression, SGDRegression, ElasticNet, AdaBoost, Ridge, RBFSVM, DecisionTree, ExtraTrees, Bagging and Random Forest, with exhaustive grid search for hyperparameters; Random Forest performed best among traditional ML algorithms. We use the prediction errors on the hold-out test set using a random 9:1 train:test split. The first three subplots represent the prediction errors using three models: Random Forest, 17-layer Plain Network and 17-layer IRNet; the last subplot contains the cumulative distribution function (CDF) of the prediction errors using the three models.