Thanks to all of the reviewers for giving the paper a thorough evaluation. Overall, the reviews struck on what we felt were the paper's strengths and weaknesses.$
This paper is an application of machine learning to the field of surface science with implications in environment and sustainability. It was written to be accessible to readers unfamiliar with Gaussian processes, as well as to those unfamiliar with surface science. It shows how to apply Bayesian optimization to a practical problem without assuming an extensive background in Bayesian statistics and machine learning, which we feel is a strength relative to other applications papers.
Although the problem of rapidly identifying adsorption sites on chemical surfaces may seem rather esoteric to the machine learning audience, it is an important problem in surface science, and we anticipate our paper having a notable impact in that field.
Specific responses for Assigned_Reviewer_1:
- Thank you for pointing out specific parts of the paper where additional clarification would be warranted.
- Discussing the generalization of our techniques would add more mathematical complexity to the paper, and as one of the other reviewers mentioned, there is enough existing literature on Bayesian optimization for readers interested in general techniques.
- An Angstrom is a unit of measurement equal to 1E-10 meters, very near the length of an atomic bond.
- By periodicity "around" the sphere, we meant that the right edge of the planar Hammer projection should be the same as the left edge. In space, they correspond to the same line of longitude on the sphere.
- CMH was mentioned earlier in the paper than DE because it is a method already documented in literature to solve the surface science problem, while DE is simply a general method for global optimization.
- The expression on line 267 is the general form of the standard periodic kernel. The x1/x2 in that expression is not the same as the x parameter from the objective function. We will make sure to clarify that point. The x/y/z axes of rotation are defined in Figure 1.
- The DFT and LJ potential energy functions have different values in Figure 7 because they make different assumptions about the system. LJ is much older (1920's) than DFT (1960's). This distinction is mentioned in the text of the paper.
- In Figures 8 and 10, as explained in the body text, BASC holds the relative positions of the nuclei fixed, while CMH lets them relax, which is why CMH has a lower potential energy even though BASC and CMH found the same global minimum. We considered this a technical detail of the problem that was difficult to express in the figure without causing more confusion.
Specific responses for Assigned_Reviewer_3:
- The paper and our approaching to solving the problem at hand are pragmatic, which we feel is a strength, and sets it apart from other papers in machine learning, which tend to be dogmatic.
- Kernel functions on a generalized Hilbert sphere could be an alternative to our approach involving great circle distances in a squared exponential kernel. It would be interesting to investigate the viability of those two approaches. We will add a citation pointing the reader to the alternate options for the spherical kernel.
Specific responses for Assigned_Reviewer_4:
- There is a small handful of Bayesian optimization applications in the much broader field of computational chemistry, such as the one you referenced, but we believe BASC to be the first application in the subfield of surface science. We will cite your reference as an example in a related field and clarify the novelty of our approach in surface science.