The understanding of natural phenomenon has been a critical in solving complex problems in computer science. In the case of statistical mechanics, scientists have applied mathematical models of particle behavior in information theory, pattern analysis and artificial intelligence. Shannon entropy which quantifies information content in messages is one of the best known measures. In machine learning, spin glass theory have been applied to optimize restricted Boltzmann machine. In this work, a new energy distribution ratio (EDR) model is proposed as a split measure for selecting the best feature in decision trees. The EDR is derived from Maxwell Boltzmann distribution where the energy of an attribute is computed as the log of ratio of sum of unique classes to the product of number of instances belonging to these classes. Thereafter, the EDR is computed as ratio of class to attribute energy. The minimum EDR value is used for selecting splitting attribute as it indicates the better distribution of energy. For testing EDR, five publicly available datasets were used for the experiment and results were compared with Information Gain and Gini-Index. The EDR split criteria provides at least 1–5 percent better accuracy, precision, recall and F1-score values for majority of the datasets.