J.C. Sch¨on, M.A.C. Wevers and M. Jansen
Understanding which compounds are capable of existence in a given chemical system, is a central goal of solid state chemistry. While the successful synthesis of these, perhaps even metastable, compounds is the final touchstone, theoretical investigations play an ever-increasing role, predicting not yet synthesized compounds, suggesting phase transitions, and yielding hints at possible synthesis routes. The key for this enterprise is the exploration of the energy landscapes of solids, which commonly exhibit a multitude of local minima and a highly complex barrier structure. Such complex landscapes are not only relevant to crystalline solids. The dynamic and static features of a large variety of systems ranging from the relaxation dynamics in glasses and spin glasses, over the folding transformation in proteins and the study of the properties of clusters and polymers, to the efficiency of combinatorial optimization algorithms, are intimately related to the structure and properties of their respective energy surfaces.
Thus, we have been developing methods to determine important global and local features of energy landscapes that are typically present in all types of complex systems, e. g., local minima and larger basins containing many minute local minima, the energetic and entropic barriers between them, the saddle points and transition regions, and the local densities of states. The algorithms we employ are based on random walks on the potential energy surface of the system. For the purpose of illustration, the landscape of CaF2(two formula units per simulation cell, Z = 2) is considered, where empirical potentials have been used for the energy evaluation.
Starting points in such an exploration are the critical points of a landscape, in particular the local minima, which in solids correspond to compounds that should be capable of exis-tence at least at low temperatures. A large number of methods exist for their determination, although, for typical complex systems, only a small fraction thereof can actually be deter-mined with a reasonable effort. However, in certain systems, e. g., the crystalline solids we are investigating, many of these minima are essentially equivalent with respect to their physical properties, e. g., differing only by very minor distortions of an ideal structure. Un-less such variations are of special interest, one collects all the corresponding local minima in one large basin or minimum region.
For the next step, the determination of the energetic barriers separating such regions, we have developed the so-called threshold algorithm. Starting from a local minimum, one performs a random walk, where every step is accepted, as long as a prescribed energy lid is not crossed. Periodically during this so-called threshold phase, quench runs (ran-dom walks, where only steps that decrease the energy are accepted) are performed. Such threshold runs are repeated for a sequence of increasing energy lids. For very low lids, the quench runs will always end in the starting minimum. The first lid value, where another
follows by using the minima as end-nodes of the tree, which are connected by internal nodes at the energy barrier between the minima (see Fig. 5). Here, the global minimum (VIII-a) corresponds to the experimentally observed structure.
-6.49
VII-c
-6.68
VII-b
-6.56
VI-f
-6.86
VIII-a
-6.46
VIII-b
-6.60
-6.90 -6.00
-6.30
-6.70 -6.50 0
E [eV/atom]
-6.40
-6.80 -5.90
-6.20 -6.10 -5.80
Figure 5: Tree-graph diagram for CaF2. Connections of lines mark lid values where a transition to other minima is found for the first time. The height of the bars on the minima indicates the energy up to which 100% (dark) and 80% (light) of the threshold runs ended at the starting minimum.
The shaded ellipse indicates the energy region, where the entropically stabilized structure ‘VII-a’ is located.
Such threshold runs involve many millions of energy evaluations. By sampling the energies during such a run, we gain information about the local density of states accessible during the random walk. By combining the samples for many lid values, we can derive accessible local densities of states for each of the minimum regions. In principle, this information can be used to give estimates of the thermodynamic properties of each local minimum region, e. g., the specific heat. Furthermore, the local density of states can also have an effect on the kinetics of the system, if the growth of the local density of states with energy is exponential with growth factor. In this case, a random walker at temperature T will not be able to leave the region for T<1/= Ttrap, while for T>1/, the walker cannot enter the region. Such approximately exponentially growing densities of states have already been observed in a number of systems ranging from spin glasses, lattice glasses and polymers to crystalline solids [ Sibani et al., Comp. Phys. Comm. 116, 17 (1999); Wevers et al., J.
Figure 6: Transition map for the major regions of the landscape of CaF2. Ar-row sizes proportional to frequency of oc-currence, ranging from 20%, 10%,
5% to<5%. Energy of minima are coded in grayscale (cf Fig. 5). Colors of arrows in-dicate height of transition above minimum:
black0.4, blue0.6, green0.8, yellow
1.2, red>1.2.
Clearly, the barrier between two minimum regions cannot be completely described by the energy of, e. g., a saddle point alone. Instead the entropic barrier, measuring the difficulty of finding a path from one region to the other, must be estimated, too. Such barriers make themselves felt in e. g., transition probabilities between two minima, which remain essen-tially zero even though it is known that a low-energy path exists. We have cast this behavior in the form of so-called return-energies that indicate the energies up to which a certain per-centage of threshold runs does not leave the starting minimum region (see Fig. 5). This is complemented by the so-called transition maps, where we represent the energy-dependent transition probability matrix in diagrammatic form (see Fig. 6).
Figure 7: Quench and low-temperature Monte-Carlo runs, exhibiting the entropically stabilized region ‘VII-a’.
Another effect is the entropic stabilization of regions of the landscape that do not corre-spond to a local minimum, but exhibit nevertheless long residing times. A beautiful ex-ample is found in the CaF2 system, where an intermediate region of the landscape, called VII-a, has been observed at the end of simulated annealing runs. A closer investigation shows that no local minimum is associated with this region. But nevertheless, Fig. 7 shows that essentially all quench runs and low-temperature (i. e. T<Ttrap (VII-a)) Monte-Carlo runs reside for a very long time in this region before reaching the global minimum VIII-a.
This result might also be of relevance regarding the question of possible intermediates in the high-pressure phase transition of RuO2from rutile over a CaCl2-intermediate to the flu-orite structure, which corresponds to the VI-f!VIII-a transition. Our results suggest that a structure like VII-a might occur as an intermediate between CaCl2 and fluorite, support-ing a proposal [ Haines et al., Phys. Rev. B48, 13344 (1993)] of a hypothetical intermediate very similar to our VII-a–structure. In addition, a very slow kinetics of the phase transfor-mation was observed, which is a first hint that the transition region might be entropically stabilized.