Microgravity surveys focus on localized small-scale subsurface targets and require higher measurement precision and enhanced data-processing resolution, making them suitable for detailed, engineering-grade investigations such as urban underground-space exploration. Microgravity anomaly inversion exhibits a distinctive capability for fine-scale delineation. This study investigates intelligent identification of microgravity anomalies using a twin-tunnel section of Chengdu Metro Line 9 as a representative underground-space case. A training dataset was first generated through forward modeling based on known tunnel parameters, and a fully connected neural network was developed to establish an end-to-end mapping from microgravity anomaly curves to tunnel geometric parameters. Synthetic experiments demonstrate that the model can accurately recover key geometric attributes—including tunnel depth, radius, spacing, and horizontal position—even under noisy conditions. For real-data preprocessing, the terrain-correction range was determined using gravity-response amplitudes associated with buildings of different scales within the urban area, while the stripping procedure for shallow-interface anomalies was formulated based on gravity-response characteristics induced by bedrock undulations beneath the surficial cover. After applying Bouguer correction and shallow-interface anomaly stripping to the raw gravity data, intelligent identification of tunnel geometry was performed, and the inversion results show strong consistency with engineering geological information, confirming the reliability and applicability of the proposed model. Overall, the method achieves high-precision characterization of shallow underground-space geometry, provides excellent inversion accuracy and strong engineering applicability, and establishes a "scientific testbed" for microgravity detection of small subsurface targets, offering a practical paradigm for the optimization and innovation of related geophysical exploration techniques.