The work reported here establishes a neural network-based methodology for failure probability prediction and quality enhancement of microengine MEMS using attribute data derived from actual measurements on microengines. Two complementary backpropagation neural networks were employed - one for failure probability prediction where microengine attributes constituted the inputs while time-to-failure statistics (mean, median and shape parameters) constituted network outputs. The second neural network was for quality enhancement through attribute refinement -- inputs were time-to-failure statistics and outputs microengine attributes. Once neural network training was complete, independent data was used to validate results. Correct prediction of failure statistics as well as determination of optimal MEMS attributes for specified failure probability levels was achieved with high confidence (0.88-0.92). Low humidity (0-10% ) for example and high microengine resonant frequency coupled with microengine operation at 0.4 of resonant frequency was found to result in median times-to-failure of at least 200 million cycles.