A practical algorithm to minimize the overall error in FEM computations

Abstract

Using the standard finite element method (FEM) to solve general partial differential equations, the round-off error is found to be proportional to $N^{\beta_{\rm R}}$, with $N$ the number of degrees of freedom (DoFs) and $\beta_{\rm R}$ a coefficient. A method which uses a few cheap numerical experiments is proposed to determine the coefficient of proportionality and $\beta_{\rm R}$ in various space dimensions and FEM packages. Using the coefficients obtained above, the strategy put forward in \cite{liu386balancing} for predicting the highest achievable accuracy $E_{\rm min}$ and the associated optimal number of DoFs $N_{\rm opt}$ for specific problems is extended to general problems. This strategy allows predicting $E_{\rm min}$ accurately for general problems, with the CPU time for obtaining the solution with the highest accuracy $E_{\rm min}$ typically reduced by 60\%--90\%.