The blind hyperspectral unmixing of altered minerals holds significant indicative importance for the exploration of sandstone-type uranium deposits. Nevertheless, the distinctive backscattering and multiphase characteristics are typically underestimated by existing methods, resulting in limited accuracy in the unmixing performance of complex mixtures. This study presents a physics-informed Swin Transformer Network (PIST–Net) for the blind hyperspectral unmixing of hematite, goethite, biotite, and chlorite. The proposed model includes a lightweight Swin Transformer encoder for the long-range modeling of spatial and spectral features. Furthermore, a dual-branch decoder with adaptive physical parameters was introduced in the spectral reconstruction section. In this decoder, the backscattering prediction head independently estimates the backscattering factor for each endmember, accounting for the true behavior of the mineral materials. As a key parameter of the Hapke model, the learnable strategy can reduce the spectral error in the single-scattering albedo (SSA) space to improve the accuracy of abundance estimation. The results show that PIST–Net consistently outperformed all six competing models on the altered mineral (AM) dataset and the NASA Reflectance Experiment Laboratory (RELAB) dataset. For mixtures composed of 2–4 endmembers, the mean RMSE of estimated abundances ranges from 0.0312 to 0.1037.

The accurate and efficient acquisition of rock mechanical properties is critical for ensuring the safety and efficiency of underground engineering construction. Traditional laboratory tests are characterized by long cycles, high costs, and an inability to reflect in situ mechanical properties, while existing deep learning models based on while-drilling data suffer from poor noise robustness, insufficient deep feature extraction, and low accuracy in synchronous multi-parameter prediction. To address these limitations, this paper proposes a hybrid deep learning model (CNN-LSTM-MoE) combining a convolutional neural network (CNN), a long short-term memory network (LSTM), and a mixture of experts (MoE) system. The model enables intelligent prediction of elastic modulus, Poisson’s ratio, and yield stress from while-drilling parameters. The proposed model integrates CNN’s local feature extraction capability, LSTM’s temporal dependency modeling capability, and the multi-expert dynamic fusion mechanism of MoE. Furthermore, it incorporates physical constraints from rock fragmentation mechanics and an adaptive multi-objective loss weight optimization strategy to comprehensively enhance the multi-parameter synchronous prediction performance. Experimental results demonstrate that the proposed model achieves coefficients of determination (R²) of 0.8965 for elastic modulus, 0.9193 for Poisson’s ratio, and 0.9813 for yield stress on the laboratory validation dataset, with a mean squared error (mse) of 4.0720. Its prediction performance significantly outperforms benchmark models such as TCN and Transformer time-series architectures. Ablation studies further validate the critical role of the integrated LSTM and MoE modules in improving model accuracy, with the MoE module contributing an average R² improvement of approximately 24%. This study not only provides an effective method for high-precision acquisition of rock mechanical parameters while drilling, but also offers a feasible solution based on numerical simulation for data augmentation to address the common issue of scarce labeled data in deep learning applications within engineering fields.