Hybrid machine learning and TOPSIS framework for domain-specific crystal selection in 2-µm Tm:YAG and Tm:LuAG lasers

Abstract

Thulium-doped garnet crystals serve divergent application domains—surgical, lidar, spectroscopic, and industrial—each imposing a distinct performance hierarchy on the laser crystal, yet no quantitative framework guides crystal selection across these regimes simultaneously. Three supervised regression models predict continuous-wave (CW) output power, Q-switched average power, and pulse energy across 20 Tm:YAG and Tm:LuAG configurations. Configuration-based GroupKFold cross-validation (n = 80) prevents data leakage in CW prediction, yielding R 2  = 0.922 with Ridge regression; leave-one-out cross-validation (n = 20) yields R 2  = 0.888 for average power and R 2  = 0.820 for pulse energy. Feature ablation confirms doping concentration as the dominant predictor (ΔR 2  =  + 0.323). TOPSIS multi-criteria ranking uses criterion weights derived from systematic analysis of 36 primary-literature sources spanning Medicine, Defence, Science, and Industry, replacing subjective expert assignment with citation-traceable weight vectors. Monte Carlo sensitivity analysis (± 20% weight perturbation, n = 500) confirms Tm:YAG 6%/5 mm as optimal for Medicine (100% stable), Defence (100% stable), and Science (70% stable), while Tm:YAG 6%/7 mm is preferred for Industry (72% stable). High Spearman rank correlations between domain rankings (ρ = 0.794–0.946, all p  < 0.001) indicate crystal physics constrains the solution space. The rankings are applicable to the criteria and literature-derived weights adopted in this study rather than universal per-domain optima, and should be interpreted with the sample size (n = 20 configurations) in mind; the framework nonetheless provides a citation-traceable, reproducible starting point for 2-µm laser crystal selection. The criterion weights are derived from the emphasis given to each performance metric across the 36 primary sources reviewed, not from a first-principles physical model of each application, and the resulting rankings are restricted to the continuous-wave and passively Q-switched Tm:YAG/Tm:LuAG configurations actually measured; they are not presented as applicable to SESAM mode-locked, self-mode-locked, single-frequency, or narrow-linewidth operation, whose governing performance criteria were not measured in this dataset.