Artificial neural networks are attracting increasing attention in various applications. They can be used as ‘universal approximations’, which substitute computationally expensive algorithms by relatively simple sequences of functions, which simulate a reaction of a set of neurons to the incoming signal. In particular, neural networks have proved to be efficient for parameterization of the computationally expensive radiative transfer models (RTMs) in atmospheric remote sensing. Although a direct substitution of RTMs by neural networks can lead to the multiple performance enhancements, such an approach has certain drawbacks, such as loss of generality, robustness issues, etc. In this regard, the neural network is usually trained for a specific application, predefined atmospheric scenarios and a given spectrometer. In this paper a new concept of neural-network based RTMs is examined, in which the neural network substitutes not the whole RTM but rather a part of it (the eigenvalue solver), thereby reducing the computational time while maintaining its generality. The explicit dependencies on geometry of observation and optical thickness of the medium are excluded from training. It is shown that although the speedup factor due to this approach is modest (around 3 times against 103 speed up factor of other approaches reported in recent papers), the resulting neural network is flexible and easy to train. It can be used for arbitrary number of atmospheric layers. Moreover, this approach can be used in conjunction with any RTMs based on the discrete ordinate method. The neural network is applied for simulations of the radiances at the top of the atmosphere in the Huggins band.
Continuous nonlinear eigenvalue solver with applications to the design of electro/magnetorheological sandwich structures
