Calibration transfer via filter learning.

Area of Science:

• Analytical Chemistry
• Spectroscopy
• Chemometrics

Background:

• Calibration transfer is crucial in analytical chemistry to prevent full recalibration.
• Existing methods like piecewise direct standardization and dynamic orthogonal projection require multiple reference samples.
• Reducing the number of reference samples is key to improving calibration transfer efficiency.

Purpose of the Study:

• To develop a new calibration transfer method requiring minimal reference samples.
• To enable efficient calibration transfer from a master instrument (source domain) to a slave instrument (target domain).
• To provide an unsupervised approach for calibration transfer, eliminating the need for target domain labels.

Main Methods:

• Utilizes a calibration database from a master instrument and a single spectrum from a slave instrument.
• Generates a source domain counterpart of the target spectrum using a multivariate Gaussian kernel.
• Trains a filter to align the response functions of the master and slave instruments.
• Implements an unsupervised strategy to avoid target domain labels.

Main Results:

• The proposed method successfully transfers calibration with only one reference sample.
• Demonstrated effectiveness on simulated and two real-world datasets.
• Outperforms several state-of-the-art calibration transfer methods.

Conclusions:

• The new method simplifies and speeds up calibration transfer by requiring only one reference sample.
• Offers a viable unsupervised calibration transfer alternative.
• Presents a promising and efficient tool for practical calibration transfer applications.