The combination of heterogeneous resources within exascale architectures guarantees to be capable of revolutionary compute for scientific applications. There will be some data about the status of the current progress of jobs, hardware and software, memory, and network resource usage. This provisional information has an irreplaceable value in learning to predict where applications may face dynamic and interactive behavior when resource failures occur. In this paper was proposed building a scalable framework that uses special performance information collected from all other sources. It will perform an analysis of HPC applications in order to develop new statistical footprints of resource usage. Besides, this framework should predict the reasons for failure and provide new capabilities to recover from application failures. We are applying HPC capabilities at exascale causes the possibility of substantial scientific unproductiveness in computational procedures. In that sense, the integration of machine learning into exascale computations is an encouraging way to obtain large performance profits and introduce an opportunity to jump a generation of simulation improvements.