TITLE: A mass spring model for hair simulation
QUESTION [14 upvotes]: A strand of hair is represented by a set of particles connected by springs.
The velocity for a particular particle is calculated implicitly using the following formula:
$\boldsymbol{v}^{n+1/2}=\boldsymbol{v}^{n}+\frac{\Delta t}{2}\boldsymbol{a}(t^{n+1/2},\boldsymbol{x}^{n},\boldsymbol{v}^{n+1/2})$
The force or acceleration ($\boldsymbol{a}$ in the above equation) produced by the spring between two adjacent particles is given by the following:
$\boldsymbol{F}^{n+1}=\frac{k}{l_{0}}\left((\boldsymbol{x}_{2}^{n}-\boldsymbol{x}_{1}^{n})^{\mathrm{T}}\hat{\boldsymbol{d}}^{n}-l_{0}\right)\hat{\boldsymbol{d}}^{n}+\Delta t\frac{k}{l_{0}}(\boldsymbol{v}_{2}^{n+1}-\boldsymbol{v}_{1}^{n+1})^{\mathrm{T}}\hat{\boldsymbol{d}}^{n}\hat{\boldsymbol{d}}^{n}$
where $\hat{d}^{n}=(x_{2}^{n}-x_{1}^{n})/\left\Vert x_{2}^{n}-x_{1}^{n}\right\Vert$
My question is, how do you numerically calculate $\boldsymbol{v}_{1}^{n+1}$ and $\boldsymbol{v}_{2}^{n+1}$ in order to calculate the force?
I've tried using Newton's Method but calculating derivative the spring force is just so complicated.
I'm attempting to implement the techniques found in this paper: link

REPLY [2 votes]: Notice that the only unknown quantities in your force equation are the updated velocities $v_i^{n+1}$, and that force depends linearly on these. Everything else is either known from the last time step, e.g. $x_i^n$, or is a simulation parameter/material constant.
Therefore, after substituting for $\mathbf{a}$ in your first equation, you will end up with a set of linear equations in the unknowns $v_i^{n+1/2}$; in particular the system will be of the form
$$(\mathbf{I} - c\Delta t^2\ \mathbf{d}\mathbf{d}^T)\mathbf{v}^{n+1/2} = \mathbf{b}$$
for some scalar $c$ and right-hand side $\mathbf{b}$. You can readily solve this system without Newton's method. In particular, for sufficiently small time steps $\Delta t$ the matrix is positive-definite, in which case I recommend using Conjugate Gradients to solve the system iteratively without ever even needing to form the dense matrix $\mathbf{d}\mathbf{d}^T.$