The precise classification of changes in rock boundaries/facies from well-log records is a complex problem in geophysical data processing. Observed well-log data are a complex superposition of nonstationary/nonlinear signals of varying wavelengths and frequencies, shaped by the heterogeneous composition and structural variation of rock types in the earth. This impairs our ability to use traditional statistical techniques, which in most cases fail to discriminate and/or, at best, do not precisely extract facies changes from complex well-log signals. We propose a new method, set in a Bayesian neural network (BNN) framework and using a powerful hybrid Monte Carlo simulation scheme to identify facies changes from complex well-log data. We first construct a complex, composite, synthetic time series using the data from three simple models: first-order autoregressive, logistic, and random white noise. Then we attempt to identify individualsignals in the pooled synthetic time series. We use the autocorrelation and the spectral characteristics of the individual signals as input vectors for training, validating, and testing the artificial neural network model. The results show that the Bayesian separation scheme provides consistently good results, with accuracy at more than 74%. When the method was tested using well-log data from the German Continental Deep Drilling Program (KTB), it was able to discriminate boundaries of lithofacies with an accuracy of approximately 92% in validation and 93% in test samples. The efficacy of the BNN in the presence of colored noise suggests that the designed network topology is robust for up to 30% correlated noise; however, adding more noise (say, 50% or more) obscures the desired signals. Our method provides a robust means for decoding finely detailed successions of lithofacies from complex well-log data, better describing the nature of the underlying inhomogeneous crust.