Accelerated segregated finite volume solvers for linear elastostatics using machine learning

The segregated solution algorithm is widely used for solving finite volume continuum mechanics problems. One major contributor to the computational time requirement of this approach is the high number of outer iterations needed to achieve convergence. The methodology proposed in this work aims to decrease the computational time required by employing an artificial neural network to predict converged solution fields for linear elastostatic finite volume analyses. The machine learning model is trained on coarse mesh data using a sequence of consecutive initial unconverged displacement fields as inputs and the converged displacement field as the target. Subsequently, the trained model is used to predict the converged displacement field for a fine mesh case. The speedup calculation incorporates the time required to run the coarse mesh case and train the machine learning model. The typical speedups achieved using the proposed technique in this study range between 2 and 4. However, it has the potential to achieve higher speedups, with the maximum observed in this study being 13.3. • A novel method for accelerating finite volume linear elastostatics segregated solvers. • Unconverged displacement fields are used to predict the converged displacement field. • The methodology produces speedups of up to 13.3 relative to the conventional approach.