1: \begin{abstract}

2: 		We implemented a geometry optimizer

3: 		based on Gaussian process regression (GPR) to find minimum structures

4: 		on potential energy surfaces.

5: 		We tested both a two times differentiable 

6: 		form of the Mat\'{e}rn kernel and the squared exponential

7: 		kernel. The Mat\'{e}rn kernel performs much better.

8: 		We give a detailed description 

9: 		of the optimization procedures. These include

10:                 overshooting the step resulting from GPR in order

11:                 to obtain a higher degree of interpolation

12:                 vs. extrapolation. In a 

13: 		benchmark against the 

14: 		L-BFGS  

15: 		optimizer of the \dl library on $26$ test systems,

16: 		we found the new optimizer to generally reduce the

17:                 number of required optimization steps. %be comparably efficient. 

18: 		%In the largest and most realistic

19: 		%test system the GPR optimizer performs 

20: 		%better than L-BFGS:

21: 		%The individual steps take longer but the number of steps

22: 		%needed to reach convergence is lower.

23: 		%In summary this yields	a faster convergence than L-BFGS.

24: 	\end{abstract}

25: