Improving Pre-Concentration Stream Fidelity in Dual-Energy XRT Ore Sorting with Process-Aware Sorting Knowledge

Sensor-based ore sorting can discard barren rock before grinding, but dual-energy X-ray transmission (DE-XRT) grade models are often judged by block-wise error that poorly reflects concentrate and tailings mass balance at the plant cut-off. Process-aware sorting knowledge (PSK) encodes attenuation, sorting tendency, and grade-band text from patch statistics and split rules; captions are aligned via a frozen contrastive language–image pre-training (CLIP) encoder during training, while inference remains Vision-only. Stream outcomes are reported as metal recovery, mass yield (mass fraction to concentrate), concentrate and tailings grades, and enrichment ratio, with the Process Fidelity Index (PFI) summarizing oracle-aligned fidelity at cut-off τ. On a fixed PbZn holdout with full assay labels, PSK raises the PFI from 0.716 to 0.830 and lowers mean absolute error (MAE) from 4.59% to 3.77% versus Vision-only training; categorical and physics-feature controls support added value from CLIP alignment. Backbone ablation shows positive PSK gains on ResNet-18, MobileNet v3 small, and ShuffleNet v2_×0.5 (ΔPFI = +0.114/+0.118/+0.089), with lightweight Vision PFI near MobileNet and below ResNet-18. Field ablation shows that attenuation and sorting-tendency fields drive most stream gains, while the grade-band field trades a higher PFI for worse tailings grade when removed.