Benchmarking Optimization of Atomic Structures on Rough Landscapes

Efficient optimization of atomic structures is crucial for accurately simulating and understanding the behavior of materials and molecules. Here, we present a comprehensive benchmarking study that evaluates the performance of various optimization methods on atomic structures, encompassing a wide range of atomic interactions. We study gradient-based methods such as Adam, SGD, second-order optimization methods including BFGS and conjugate gradient, molecular dynamic-based techniques like FIRE, Basin Hopping, and simulated annealing, as well as neural methods such as learned optimizers and reinforcement learning-based optimizers. Chosen systems represent diverse atomic interactions, covering various types of potentials, including pairwise potentials such as Lennard-Jones (LJ), three-body angular interactions represented by the Stillinger-Weber potential, Coulombic interactions in the Buckingham potential, and multi-body interactions in the embedded atom potential. We evaluate the performance of each optimization method on different systems, measuring factors such as final energy, computational efficiency, and scalability. Neural techniques leverage the power of neural networks to adaptively optimize the structures and capture complex relationships between atomic interactions. Initial findings suggest that neural methods hold great potential for enhancing the optimization process and achieving higher accuracy in complex atomic systems.