Identification of saddle points on potential energy surfaces (PES) is of great importance in materials science, computational chemistry and related fields. Indeed, the long-term dynamics of a system is determined by the saddle points of its PES. To this end, numerous algorithms have been, or are currently being developed. These algorithms can be roughly divided into two types, (i) double-ended methods, which require prior knowledge of both initial and final state of the process under study, and (ii) single-ended methods, which allow for the exploration of the PES from a single initial configuration. One of the broadly adopted double-ended methods is the Nudged-Elastic-Band (NEB) [1], developed by Jónsson, whereas examples of single-ended methods include the Activation-Relaxation Technique (ART), developed by N. Mousseau in 1996 [2], and the dimer method, developed by Jónsson in 1999 [3]. These methods have been improved since their initial conception, and novel approaches for identifying saddle points are also emerging [4, 5, 6, 7, 8]. The success of these approaches is clearly demonstrated by the number of citations of reference papers (NEB 14k, Dimer 3k, ART 1k), as well as by their implementation in most materials modelling legacy codes (Quantum ESPRESSO, VASP, LAMMPS).

With the increased need to accurately and efficiently address the dynamical phenomena in more and more realistic systems, it is now the right time to bring together the developers and contributors of the main saddle search algorithms, and to define an extensive common benchmark framework as to be comparable to those already established in other areas of computational materials science, i.e., the Standard Solid-State Pseudopotentials (SSSP) library [9], and GW100 [10]. First steps in this direction have already been taken (OptBench, [11]). OptBench offers a series of well-defined tests, each of them testing a different segment or purpose of the software. Among others, the tests include a general exploration of the PES for saddle points starting from an initial minimum, as well as more particular and targeted tests, such as saddle point refinement, and performance in determining the lowest eigenvector. Depending on the purpose of a test, there are several criteria for the performance of the software in that test, with the main focus being on the number of force calls needed to complete the given test. Currently, the available tests are limited to empirical interatomic potentials, on Lennard-Jones clusters and surfaces. The objectives of SSbench are therefore to i) discuss the possible extension of the benchmarks to molecular reactions, reactions on surfaces and diffusion processes in bulk materials, for both empirical and first-principles energy and force engines; ii) develop meaningful criteria to assess the performance of saddle search algorithms (i.e accuracy, speed, robustness and versatility) in a transparent and reproducible way; iii) discuss possible routes to make the data easily accessible; iv) discuss common strategies to develop common workflows and analysis tools.
 

SSbench will give the opportunity for developers and contributors to present their views on how to develop such a framework through a short and targeted list of invited contributions and free-format round table discussions. This workshop proposal is the outcome of a community of developers willing to join efforts to develop a common benchmark framework.