APTNN
A neural network method for solving high-dimensional kinetic equations that combines the CP decomposition structure of tensor neural networks with the physical constraints of asymptotic-preserving schemes, allowing high-dimensional loss integrals to be decomposed into one-dimensional integral computations and thereby reducing computational complexity from exponential to polynomial.
Timeline (last 90 days)
APTNN (Asymptotic-Preserving Tensor Neural Network) combines tensor decomposition with physics-constraint loss to compress the training time of the 3D3V Vlasov–Poisson–Fokker–Planck problem to under half an hour on a single GPU.
APTNN introduces the canonical polynomial (CP) decomposition structure of tensor neural networks (TNN) into the approximate representation of the solution, and the equation residual retains the CP decomposition form under this representation.
APTNN additionally adds the first-order moment equation to the loss function to constrain the limit flux closure, thereby reducing the approximation error in the high-field region.
The enhanced AP loss can still be evaluated separately by dimension, without breaking the efficiency advantage brought by the CP decomposition.
Numerical results show that APTNN achieves a marked improvement in prediction accuracy in the high-field region, while computational cost and memory usage drop substantially.
The specific example given in the paper is a 3D3V VPFP problem with multiple random inputs, whose training takes less than 30 minutes on a single GPU.
The paper does not yet provide a systematic efficiency comparison with representative conventional solvers at equal accuracy; a complete benchmark is left to future work.
APTNN does not rely on sampling tricks to approximate high-dimensional integrals; instead, it uses tensor decomposition to make the integral itself exactly decomposable and computable, which is structurally different from the statistical approximation of PINNs.
Traditional methods for uncertainty quantification (UQ) require rerunning the discrete solver for each sample, with computational cost proportional to the number of samples; APTNN's CP decomposition structure allows the network to accept random inputs as additional dimensions, so that a parametric solution mapping is obtained with a single training run.
The choice of CP decomposition rank, the weighting balance of the augmented loss terms across different parameter regions, and extension to more complex geometries and boundary conditions are questions that have not yet been fully explored.
All Facts (10)
APTNN (Asymptotic-Preserving Tensor Neural Network) combines tensor decomposition with physics-constraint loss to compress the training time of the 3D3V Vlasov–Poisson–Fokker–Planck problem to under half an hour on a single GPU.
50%UnverifiedAPTNN introduces the canonical polynomial (CP) decomposition structure of tensor neural networks (TNN) into the approximate representation of the solution, and the equation residual retains the CP decomposition form under this representation.
50%UnverifiedAPTNN additionally adds the first-order moment equation to the loss function to constrain the limit flux closure, thereby reducing the approximation error in the high-field region.
50%UnverifiedThe enhanced AP loss can still be evaluated separately by dimension, without breaking the efficiency advantage brought by the CP decomposition.
50%UnverifiedNumerical results show that APTNN achieves a marked improvement in prediction accuracy in the high-field region, while computational cost and memory usage drop substantially.
50%UnverifiedThe specific example given in the paper is a 3D3V VPFP problem with multiple random inputs, whose training takes less than 30 minutes on a single GPU.
50%UnverifiedThe paper does not yet provide a systematic efficiency comparison with representative conventional solvers at equal accuracy; a complete benchmark is left to future work.
50%UnverifiedAPTNN does not rely on sampling tricks to approximate high-dimensional integrals; instead, it uses tensor decomposition to make the integral itself exactly decomposable and computable, which is structurally different from the statistical approximation of PINNs.
50%UnverifiedTraditional methods for uncertainty quantification (UQ) require rerunning the discrete solver for each sample, with computational cost proportional to the number of samples; APTNN's CP decomposition structure allows the network to accept random inputs as additional dimensions, so that a parametric solution mapping is obtained with a single training run.
50%UnverifiedThe choice of CP decomposition rank, the weighting balance of the augmented loss terms across different parameter regions, and extension to more complex geometries and boundary conditions are questions that have not yet been fully explored.
50%