Knowing which variables affect the are most the perovskite solar cell stability could have a significant impact on the commercial transition of this technology. With this information, time and resources can be saved during synthesis by allowing the materials scientist to act on those variables that most influence in the long-time stability. Machine learning techniques (ML) can be used for this purpose. There are two approaches: a) using feature subset selection techniques and b) using statistical association measures. In the first case, the selected variables and the relevance measure they supply depend on the type of ML model used. That is, the result depends on the type and complexity of the model within the range of possibilities offered by the ML field. In contrast, statistical association measures supply a single result because they do not incorporate the use of models in their procedures. Their results are independent of the model.

Among these statistical association measures, Pearson correlation coefficient and Kendall and Spearman rank correlation measures can be mentioned. In the case of Pearson, it is assumed that the relationship between the involved random variables is linear, which is rare in problems of the complexity degree of the materials science field. On the other hand, rank correlation measures work for nonlinear phenomena, but only if the underlying function of the phenomenon is monotonic increasing or decreasing; however, this is information that is not available. In contrast to earlier measures, mutual information (MI) measure does not suffer from linearity and monotonicity restrictions.

In this work, the mutual information measure is proposed to quantify the amount of information contained in the descriptor variables of the synthesis process with respect to the long-time stability of perovskite solar cells. The stability measures correspond to the time in hours it takes for a cell to lose 5% and 20% respectively. The mutual information measure (amount of information that the descriptor contributes to the stabilities parameters) is estimated from a data set composed of over 40,000 experimental observations of perovskite cells.