MEMS time-to-failure is a random variable with a known probability density function (pdf). The approach here is to use a neural network to predict the pdf of failure time rather than a specific random failure time. A database of microengine attribute data (fabrication process, physical specifications, operating environment) and microengine performance data (time-to-failure) derived from actual measurements on fabricated microengines is employed. Two separate but complementary backpropagation neural networks are applied to data to produce accurate mappings first between microengine attributes and microengine time-to-failure and secondly between microengine time-to-failure and microengine attributes. The first neural network is for failure probability prediction, microengine attributes constituted the inputs while time-to-failure statistics (mean, median and shape parameters) constituted neural network outputs. The second neural network is for quality enhancement through attribute refinement – inputs are time-to-failure statistics and outputs are microengine attributes. Once neural network training was complete, in the first case, independent validation data was used to verify predicted microengine failure statistics and attributes. Results indicated correct prediction of failure statistics with minimum correlation coefficient of 0.95. In the second case, by reversing neural net inputs and outputs, optimal MEMS attributes for specified failure probability levels were determined also with good correlation (0.88-0.92).