Accurate reconstruction and prediction of the hohlraum x-ray drive remain major challenges in indirect-drive inertial confinement fusion, owing to the complexity of laser-to-x-ray conversion and the limited fidelity of existing radiation-hydrodynamic models. In this work, we investigate double-shock implosion experiments on a 100 kJ-class laser facility and develop a data-driven framework for reconstructing and predicting the hohlraum x-ray drive using multiple diagnostics. Using the radiation-hydrodynamics code Icefire-1D, we analyze 48 shots with keyhole and gas-filled hohlraums constrained by VISAR, neutron streak camera, and other diagnostic measurements. To improve agreement between simulations and measurements, nine drive tuning factors and one M-band fraction are introduced to jointly correct the drive history and spectral distribution in Icefire-1D. Based on the reconstructed results, a neural-network surrogate model is developed to map target design parameters to the corresponding tuning factors, enabling rapid prediction of x-ray drive waveforms from target design inputs. Building on this approach, we establish a deep-learning framework, PRISM, for pre-shot x-ray drive prediction. The model achieves high predictive accuracy and computational efficiency, making it suitable for pre-shot simulation and design studies. Its predictive performance beyond the training set is validated against transmission grating spectrometer measurements, showing good agreement between predicted and measured spectra. In addition, SHAP analysis is used to quantify the effects of design parameters on the inferred tuning factors. Residual analysis further indicates a marked improvement in shot-to-shot control and reproducibility of the 100 kJ-class facility in 2025, particularly in laser output stability, although pulse synchronization and timing control still require further improvement. This work provides a practical basis for high-fidelity x-ray drive modeling and more predictive simulations of indirect-drive ICF experiments.

laser fusion,machine learning,drive deficit