This paper proposes an intelligent prediction method for backfilled foundation settlement in loess gorge terrain, which is based on optimizing numerical simulation parameters using intelligent algorithms. The method combines experimental data, field monitoring data, and numerical simulation techniques, employing convolutional neural networks to optimize and calibrate the model parameters, thus improving the accuracy and efficiency of settlement predictions. The research results indicate that the settlement of foundations in loess gorge areas is influenced by multiple factors, particularly the compaction of backfill, the engineering properties of the original soil, and hydrological conditions. These issues must be addressed through reasonable parameter optimization and precise prediction. Traditional settlement prediction methods are less applicable in complex geological conditions and fail to effectively reflect the unique characteristics of loess gorge terrain. By integrating experimental data and field monitoring data, and optimizing numerical simulation parameters with intelligent algorithms, the prediction accuracy of settlement can be significantly improved. The introduction of intelligent algorithms allows the model to automatically adjust parameters, reducing human errors and enhancing prediction efficiency. The results of the intelligent prediction method align closely with the measured data, with errors controlled within 5%, meeting the engineering requirements. The intelligent prediction method proposed in this paper can effectively address the settlement issues of backfilled foundations in loess gorge areas, providing a valuable reference for foundation engineering in similar regions and offering technical support for the design and construction of similar projects in the future.