Rare Earth Researchers Fuse Laser Spectroscopy and AI to Improve Ore Analysis

Highlights

• A CNN-Transformer fusion model outperformed simpler deep-learning approaches in 13 of 16 single-matrix ore analysis tasks.
• The model achieved R²=0.9963 and RMSE=0.418 wt% for ytterbium under mixed-matrix conditions, a strong analytical result.
• Interpretability tools like Grad-CAM and SHAP confirmed the AI tracked genuine elemental signals rather than memorizing rock-type differences.
• Faster LIBS analysis could support exploration, ore sorting, and beneficiation closer to the mine face, but independent mine-scale validation is still needed.

A Chinese research team has given an old analytical tool a smarter brain. Lead author Huihui Zhu, corresponding author Tao Lü, and collaborators associated with China University of Geosciences (opens in a new tab) and partner institutions report (opens in a new tab) in Microchemical Journal that combining laser-induced breakdown spectroscopy (LIBS) with a CNN-Transformer artificial-intelligence model can more reliably measure rare earth elements in complex polymetallic ores. Their model beat simpler deep-learning approaches in most tests and achieved R²=0.9963 for ytterbium (Yb) under mixed-matrix conditions. For miners, the potential prize is straightforward: faster identification of what is actually inside complicated ore—without waiting for every sample to travel through a conventional laboratory.

REEx Insight: Making the Ore Talk Faster

LIBS fires a laser pulse at rock, creating a tiny plasma whose emitted light provides an elemental fingerprint. The problem is the matrix effect: iron, calcium, silicon, and other minerals can alter that signal, while rare-earth spectral lines themselves overlap densely. Zhu and colleagues attacked that problem computationally. A convolutional neural network (CNN) identifies nearby spectral patterns while a Transformer searches relationships across distant wavelengths. Importantly, the researchers also used Grad-CAM, SHAP, and other interpretability techniques to test whether the AI was following genuine elemental signals rather than simply memorizing differences between rock types.

From Laser Flash to Ore Grade

Across 16 single-matrix tasks involving four matrices, the fusion model recorded the lowest root-mean-square error in 13, while producing narrower error variation between matrices. For Yb under mixed-matrix conditions, it reached R²=0.9963 and RMSE=0.418 wt%.

That matters because faster, more reliable LIBS could eventually support exploration, ore sorting, and beneficiation, potentially pushing analytical intelligence closer to the mine face.

But this is not yet an autonomous ore-grading revolution. The reported performance comes from the researchers' experimental sample system, not independent mine-scale validation. Different deposits, mineralogy, surface conditions, instruments, and calibration populations could degrade performance. Related work from this research group also underscores that rough surfaces themselves can materially complicate LIBS measurements.

REEx Bottom Line

The breakthrough is less the laser than the ability to separate elemental signal from geological noise. Next should come blind validation across independent deposits, laboratories, and operating conditions. If those results hold, AI-enhanced LIBS could become another tool for accelerating the upstream rare-earth supply chain—but the paper demonstrates a promising analytical method, not commercial readiness.

REEx Connect

Citation: Zhu H, Wang J, Zhang Y, et al. Robust and interpretable laser-induced breakdown spectroscopy quantification of rare earth elements in polymetallic ores. Microchemical Journal (2026). DOI: 10.1016/j.microc.2026.119461. Related publications confirm this research group's continuing work on deep-learning-assisted LIBS analysis of rare-earth-bearing polymetallic ores.