Linear scaling approach

The system size problem necessitates mixed QM/MM approaches which in the future might be accompanied by linear scaling approaches. However, the most severe of the remaining limitations is the time scale of a few tens of picoseconds during which the system can be sampled. Therefore, the combination of AIMD and Hybrid/AIMD simulations with enhanced sampling techniques [163] can be expected to multiply the power of this approach. 

Often, the bottleneck in linear-scaling density-functional theory is the evaluation of the Coulomb potential the trade olf between the simple and direct method of integrating Eq. (94) and the more sophisticated linear-scaling approaches is evidenced by the fact that, for moderately large systems, linear-scaling density-functional techniques are often less efficient than direct solution to the Kohn-Sham system. As the size of the system increases beyond 10 to 20 A, however, linear-scaling techniques become essential. 

The COSMO method is also interesting as the basis of a very successful COSMO-RS method, which extends the treatment to solvents other than water [27,28]. The COSMO method is very popular in quantum chemical computations of solvation effects. For example, 29 papers using COSMO calculations were published in 2001. However, we are not aware of its use together with MM force fields. Compared with the BE method, COSMO introduces one more simplification, that of Eq. (22). On the other hand, the matrix A in Eq. (21) is positively defined [25], which makes solution of the system of linear equations simpler and faster. Also, because both A and B matrices contain only electrostatic potential terms, their computation in quantum chemistry is easier than calculation of the electric field terms in Eq. (12). Another potential benefit is that the long-range electrostatic potential contribution is easier to expand into multipoles than the electric field needed in BE methods, which may benefit linear-scaling approaches. 

This review of semiempirical quantum-chemical methods outlines their development over the past 40 years. After a survey of the established methods such as MNDO, AMI, and PM3, recent methodological advances are described including the development of improved semiempirical models, new general-purpose and special-purpose parametriza-tions, and linear scaling approaches. Selected recent applications are presented covering examples from biochemistry, medicinal chemistry, and nanochemistry as well as direct reaction dynamics and electronically excited states. The concluding remarks address the current and future role of semiempirical methods in computational chemistry. 

The MM/PCM simulation in Fig. 11.9 consists of 124,000 surface discretization points (those for which F, > 10 ). As such, solution of Eq. (11.22) by matrix inversion or other 0[N ) methods is clearly infeasible, and a linear-scaling approach (in both memory and CPU... 

Hopefully, the systematic improvements in XC functionals will indeed come about, and efficient codes will be developed to exploit such exciting new linear-scaling approaches for performing DFT calculations on large biomolecular systems. Ultimately, we could use these DFT methods to accurately and realistically model reactions in condensed media. DFT studies of enzymatic reaaions will undoubtedly yield valuable insights into often poorly understood mechanisms. Such applications may be overly ambitious and may not even become remotely feasible until well past the turn of the century. However, it is safe to say that at the present time, such goals are far more realistic, and hr closer, within DFT than within any wavefunction-based ab initio approach. 

In most QM algorithms, there will be two or more steps in a calculation with unfavourable scaling behaviour. This means that for an algorithm which scales linearly overall with system size, each separate step in the calculation must be made to scale linearly as well. Owing to the variety of different QM calculations no attempt will be made here to comprehensively list the many alternative linear-scaling approaches. Instead, we confine ourselves to a number of general remarks. 

The geometry optimizations of molecular structures were performed using the semi-empirical PM3 Hamiltonian [67,68] and the B3LYP/6-3 lG(d) method. All stationary points were confirmed to be true minima by evaluation of hessian. In order to speed up HE and DFT calculations, the fast multipole method (EMM) [47, 48, 69] has been used as implemented in Gaussian suite of programs [56]. We also used linear scaling approaches for calculations of nonlinear optical properties as implemented in ADF package [58]. All the properties are expressed in atomic units. Conversion factors can be found elsewhere [28-31],... 

---