UCT-LMIR: A standard-free unsupervised calibration transfer method without model recalibration

Cross-instrument spectral variations in intensity, baseline, and noise structure severely undermine the transferability of multivariate calibration models, limiting their practical deployment across spectroscopic platforms. Existing calibration transfer methods typically rely on standard samples, target domain reference labels, or model recalibration, all of which incur substantial experimental or computational costs.

This work proposes unsupervised calibration transfer via likelihood maximization inverse regression (UCT-LMIR), which simultaneously eliminates the need for standard samples, target domain reference labels, and model recalibration. UCT-LMIR constructs an error covariance model within a latent space and performs prediction by maximizing the likelihood of a target spectrum under the learned error structure. An optional covariance compensation term accounts for the additional dispersion induced by cross-domain spectral discrepancies, enabling the method to accommodate varying degrees of instrumental mismatch. Validation on three spectroscopic datasets with distinct distributional characteristics showed that UCT-LMIR outperformed representative unsupervised calibration transfer methods in the majority of transfer tasks, with particularly pronounced improvements under large inter-instrument distribution shifts where competing methods often failed.

UCT-LMIR provides a standard-free and unsupervised calibration transfer strategy that does not require model recalibration. By combining physical interpretability with ease of implementation, it offers a practical and robust solution for calibration transfer in real-world spectroscopic applications.