This work presents a novel methodology for quantifying laser pointing offsets to evaluate and optimize laser-target coupling in experimental configurations that lack direct line-of-sight diagnostics. To overcome this measurement limitation, we have developed a physics-informed reconstruction algorithm that utilizes multi-view X-ray imaging to establish a robust mapping between radiation asymmetries [1] and laser pointing offsets (i.e. deviations from the nominal aiming point), thereby enabling substantial improvements in the calibration fidelity and overall accuracy of laser alignment.

The study focuses on the analysis of a comprehensive dataset comprising 2,790 3D radiation-hydrodynamic simulations conducted with the state-of-the-art MULTI-3D code[2]. By re-calibrating the radiation transport calculations, specifically the mean free path and absorption rate, the simulations were able to reproduce with high fidelity 2D projection images consistent with measurements from a six-view X-ray diagnostic system. Through detailed statistical analysis, we identified consistent physical correlations: the deviation angle of the "brightness centroid" emerges as one of the most sensitive and effective parameters for quantifying the offsets, whereas the upper-to-lower brightness ratio serves as a reliable secondary diagnostic for inferring the offset along the z-direction.

Building on these physics-derived anchors, we introduce SAFIR (Soft–hard Angle Fusion for ICF offset Regression), a physics-informed, two-stage multi-view regressor. In contrast to deep convolutional neural networks (CNNs), SAFIR is a lightweight, parameterized linear regression framework (~30 trainable parameters) optimized via PyTorch. The method innovatively fuses "hard-threshold" geometric angles with "soft-dipole" features—which compute threshold-free vector resultant angles —to jointly extract robust directional features and to yield spatial confidence scores. In the first stage, the model prioritizes diagnostic equipments located in the equatorial plane to infer the offset along the y-axis. In the second stage, it estimates the z-axis offset conditioned on both the multi-axial brightness features and the previously inferred y-offset, utilizing teacher forcing during training and a self-consistent feedback pathway during inference. Furthermore, SAFIR incorporates confidence-weighted, axis-weighted, and group-consistency loss functions to address heterogeneous data distribution and temporal drift.

Quantitative evaluations demonstrate the model's exceptional accuracy and robustness across multiple temporal frames, as evidenced by the near-unity parity slopes and low-bias residuals (means around 2 μm). The primary innovation of this work lies in the deep integration of physical intuition into a highly interpretable, parameter-efficient framework. This approach not only facilitates the extraction of critical metrics of laser-target interactions from complex multi-view data, but also enables a rapid, deployable solution for high-precision laser alignment calibration in inertial confinement fusion experiments.
 

ICF diagnose,3D x-ray imaging,physical informed,teacher forcing