This work focuses on the usage and performance of convergent non-uniform asynchronous elementary cellular automata (NAECAs) for supervised machine learning tasks like pattern classification. This work targets on some theoretical findings like i) how rule min terms (RMTs) can be used to generate strictly convergent NAECAs, ii) behavior of cellular automata obtained by mixing different groups of (146) convergent AECA rules based on the different proportions of passive RMTs and iii) how the properties like number of fixed point attractors (FPAs), rate of convergence change on adding impurity to the obtained strictly non-uniform AECA. Based on the result of our proposed three strategies, potential candidate NAECA rules for pattern classification are chosen keeping constraints like average convergence steps, number of FPAs. The result of our NAECA based classifier is compared with other cellular automata and machine learning based pattern classification benchmark algorithms on real datasets.