We present a forward-modeling-based sampling of diffusion-tensor imaging (DTI) integral curves. This work has the potential to generate accurate brain neural fiber models that fit the data well with an economic number of curves. DTI integral curves are integrated from the first eigenvector field of the DTI field. Usually the seed points are generated randomly or from a regular grid in the data volume. The resulting set of integral curves is dense around the long and skinny neural fiber structures and sparse around the short and fat structures. There is currently a lack of quantitative indication of how well various models fit the data. We build a forward model that simulates diffusion-weighted images (DWIs) from the DTI integral curves based on a multi-tensor model. We employ the sum of the difference between the simulated DWIs and the acquired DWIs as the goal function and optimize the placement of the DTI integral curves with a greedy algorithm and a simulated annealing algorithm. The results show that with the same number of curves, the optimized set of DTI integral curves fit better to the data than randomly seeded integral curves.