Work

Physical consistency and completeness in the data-driven modeling of microwave absorbing materials

Year: 2026
Type: preprint
Abstract: Machine learning (ML) has been applied to the modeling of microwave absorbing materials, including the prediction of electromagnetic parameters and reflection loss (RL), yet the physical consistency of such models is rarely examined. This paper addresses two types of physical consistency problems in data-driven modeling: the physical inconsistency of predicted outputs and the physical incompleteness of input features. On the output side, we constructed a numerical integrator based on the Kramers-Kronig (KK) relation to test permittivity spectra output by ML models. The results show that none of the four examined spectra from three studies satisfied the KK relation, indicating that the model outputs markedly violate physical consistency. On the input side, we took a published MXene study as a case example. There is no thickness, which determines the RL jointly with other variables, in its feature space. Using physically consistent synthetic data, we conducted controlled experiments under two data settings and two partition schemes. It was found that without the thickness, the model captured only a limited part of the RL variance and could not distinguish the labels corresponding to different thicknesses. Adding the thickness restored the performance, with the R2 rising from 0.364 to 0.956 under the fullthickness setting. Finally, we further discussed three approaches to embedding the KK constraint into the modeling workflow. Our results suggest that physical consistency and completeness are necessary conditions for building reliable data-driven models, and that global integral constraints such as the KK relation cannot be learned implicitly from small samples.
Source: ChemRxiv
Author: Hongtao Zeng
Language: English
Open Access status: green
OpenAlex ID: W7225986429
Cites: 34
Cited by:
Related to: 10