Iteration space

These vectors are orthonormalized with respect to the initial set Y and are added to enlarge the iterational space. 

A novel optimization approach based on the Newton-Kantorovich iterative scheme applied to the Riccati equation describing the reflection from the inhomogeneous half-space was proposed recently [7]. The method works well with complicated highly contrasted dielectric profiles and retains stability with respect to the noise in the input data. However, this algorithm like others needs the measurement data to be given in a broad frequency band. In this work, the method is improved to be valid for the input data obtained in an essentially restricted frequency band, i.e. when both low and high frequency data are not available. This... 

Another approach involves starting with an initial wavefimction Iq, represented on a grid, then generating // /q, and consider that tiiis, after orthogonalization to Jq, defines a new state vector. Successive applications //can now be used to define an orthogonal set of vectors which defines as a Krylov space via the iteration (n = 0,.. ., A)... 

The size of the move at each iteration is governed by the maximum displacement, Sr ax This is an adjustable parameter whose value is usually chosen so that approximately 50/i of the trial moves are accepted. If the maximum displacement is too small then mam moves will be accepted hut the states will be very similar and the phase space will onb he explored very slowly. Too large a value of Sr,, x and many trial moves will be rejectee because they lead to unfavourable overlaps. The maximum displacement can be adjuster automatically while the program is running to achieve the desired acceptance ratio bi keeping a running score of the proportion of moves that are accepted. Every so often thi maximum displacement is then scaled by a few percent if too many moves have beei accepted then the maximum displacement is increased too few and is reduced. 

As an alternative to the random selection of particles it is possible to move the atom sequentially (this requires one fewer call to the random number generator per iteration) Alternatively, several atoms can be moved at once if an appropriate value for the maximun displacement is chosen then this may enable phase space to he covered more efficiently. 

This algorithm alternates between the electronic structure problem and the nuclear motion It turns out that to generate an accurate nuclear trajectory using this decoupled algoritlun th electrons must be fuUy relaxed to the ground state at each iteration, in contrast to Ihe Car-Pairinello approach, where some error is tolerated. This need for very accurate basis se coefficients means that the minimum in the space of the coefficients must be located ver accurately, which can be computationally very expensive. However, conjugate gradient rninimisation is found to be an effective way to find this minimum, especially if informatioi from previous steps is incorporated [Payne et cd. 1992]. This reduces the number of minimi sation steps required to locate accurately the best set of basis set coefficients. 

FIGURE 21.3 Sampling of confonnation space using a Monte Carlo search (with a small number of iterations). 

Once a direction is estabflshed for the next poiat ia the space of the variables of optimization (whether by random search, by systematic evaluation of gradients, or by any other methods of making perturbations), it is possible to take a jump ia the directioa of the improvement much greater than the size of the perturbations. This could speed up the process of finding the optimum and reduce computer time. If such a leap is successful, the next iteration may take a bigger leap and so on, until the improvement stops. Then one could reverse the direction and decrease the size of the step until the optimum is found. 

In his early survey of computer experiments in materials science , Beeler (1970), in the book chapter already cited, divides such experiments into four categories. One is the Monte Carlo approach. The second is the dynamic approach (today usually named molecular dynamics), in which a finite system of N particles (usually atoms) is treated by setting up 3A equations of motion which are coupled through an assumed two-body potential, and the set of 3A differential equations is then solved numerically on a computer to give the space trajectories and velocities of all particles as function of successive time steps. The third is what Beeler called the variational approach, used to establish equilibrium configurations of atoms in (for instance) a crystal dislocation and also to establish what happens to the atoms when the defect moves each atom is moved in turn, one at a time, in a self-consistent iterative process, until the total energy of the system is minimised. The fourth category of computer experiment is what Beeler called a pattern development... 

Molecular dynamics, in contrast to MC simulations, is a typical model in which hydrodynamic effects are incorporated in the behavior of polymer solutions and may be properly accounted for. In the so-called nonequilibrium molecular dynamics method [54], Newton s equations of a (classical) many-particle problem are iteratively solved whereby quantities of both macroscopic and microscopic interest are expressed in terms of the configurational quantities such as the space coordinates or velocities of all particles. In addition, shear flow may be imposed by the homogeneous shear flow algorithm of Evans [56]. 

---