Optimizing shale lithofacies classification through advanced intelligent models in Hongxing Area, Southwest China

A "three-endmember-four-component" classification scheme, based on mineral composition and organic matter content, is adopted to divide the shale lithofacies in the Hongxing area into five categories: carbon-rich mixed, carbon-rich siliceous, high-carbon mixed, high-carbon siliceous, and medium-carbon calcareous. Bayesian classification, BP neural networks, and random forest methods are used to identify lithofacies. Bayesian classification performs better in learning lithofacies classification patterns with limited samples but is less robust to sample errors. BP neural networks excel at handling complex nonlinear relationships and achieve higher accuracy under ideal conditions but are highly dependent on sample size and model parameters. Random forest provides stability through ensemble decision trees but is less effective than BP neural networks in learning complex nonlinear lithological relationships. To enhance lithofacies identification, a heterogeneous ensemble learning model integrating the strengths of neural networks, random forest, and Bayesian classification is proposed and applied to lithofacies prediction in the Hongxing area.