The content discusses the use of Invertible Neural Networks (INNs) to reconstruct the bubble distribution in water electrolysis cells from external magnetic field measurements. The key points are:
Electrolysis for hydrogen production is hindered by gas bubbles that form during the process, reducing cell efficiency and increasing energy consumption. Detecting the bubble size and distribution is crucial for improving the process.
An approach called Contactless Inductive Flow Tomography (CIFT) can reconstruct flow fields in conducting fluids by measuring induced electric and magnetic fields. However, in the current setup, the limited number of sensors and the lack of an externally applied magnetic field pose challenges in achieving satisfactory reconstruction of the high-dimensional current distribution.
The authors explore the use of INNs to reconstruct the high-resolution conductivity map from the low-resolution magnetic field measurements. INNs can learn a bijective mapping between the conductivity map and the combination of magnetic field measurements and latent variables, addressing the information loss in the forward process.
The authors evaluate the INN-based approach against classical regularization techniques like Tikhonov and ElasticNet. The results show that the INN can effectively reconstruct the conductivity map and the location of non-conductive bubble fractions, even with a limited number of sensors and increased sensor distance.
To quantitatively evaluate the performance, the authors use a randomized error diffusion technique to convert the continuous conductivity maps into binary ensembles. The likelihood of the groundtruth with respect to the estimated probability distribution of the binary ensembles is used as the evaluation metric, confirming the superior performance of the INN model.
The authors conclude that INNs offer a promising approach for localizing and estimating non-conductive fractions in current-conducting liquids, with potential for practical applications in water electrolysis.
다른 언어로
소스 콘텐츠 기반
arxiv.org
더 깊은 질문