Calibration is a key part of the development of a diagnostic. Standard approaches require the setting up of dedicated experiments under controlled conditions in order to find the calibration function that allows one to evaluate the desired information from the raw measurements. Sometimes, such controlled experiments are not possible to perform, and alternative approaches are required. Most of them aim at extracting information by looking at the theoretical expectations, requiring a lot of dedicated work and usually involving that the outputs are extremely dependent on some external factors, such as the scientist experience. This work presents a possible methodology to calibrate data or, more generally, to extract the information from the raw measurements by using a new unsupervised physics-informed deep learning methodology. The algorithm allows to automatically process the data and evaluate the searched information without the need for a supervised training by looking at the theoretical expectations. The method is examined in synthetic cases with increasing difficulties to test its potentialities, and it has been found that such an approach can also be used in very complex behaviours, where human-drive results may have huge uncertainties. Moreover, also an experimental test has been performed to validate its capabilities, but also highlight the limits of this method, which, of course, requires particular attention and a good knowledge of the analysed phenomena. The results are extremely interesting, and this methodology is believed to be applied to several cases where classic calibration and supervised approaches are not accessible.
© 2023. The Author(s).