Automatic differentiation memory footprint

The memory footprint of AD is too high for practical unsteady cases: recording the unsteady solution for one mode and one frequency of a wing counting about 2000 panels requires about 60GB of memory. This is probably mainly due to the fact that panel methods scale quadratically with the number of panels. Two solutions have been tried:

• reduce the number of panels in the wake;
• use farfield approximations for calculating the AIC.

The first solution did not work at all and it was impossible to reduce the number of panels in the wake while keeping the accuracy of the flow solution. The second solution yielded a decrease of 20-30% of the computational time for an approximated, but still valid solution. However, the memory footprint was still the same.

In order to circumvent this issue, the best solution is probably to implement a fast multipole method, or another type of clustering method. This would change the complexity of the panel method and reduce botht the computational cost and memory. However, it requires major changes in the code.