Open Research Europe presents Michele Ceriotti, one of the authors of ‘Local invertibility and sensitivity of atomic structure-feature mappings’, in this month’s research spotlight. Read on to find out more about the research.

Can you briefly tell us about your research? 

Machine learning techniques have become commonplace in atomic-scale modeling, where they replace computationally demanding quantum mechanical calculations. However, many fundamental aspects of their use are still not fully understood. The first step in applying machine learning techniques to atomistic simulations is to map the coordinates of the atoms into a description that is more symmetric e.g. that is independent of the position or the orientation of a molecule in space. The properties of such mappings e.g. how much the symmetric descriptors change upon deformations of the original structure, have an effect on any model built on top of them. In this study, we investigate the properties of the most widely used families of descriptors, revealing how both their fundamental nature and the details of how implementing them affect the capacity of the descriptors to respond to a molecular distortion and the accuracy of the subsequent machine learning models.

We show that symmetric structures are bound to have deformation patterns that do not change the value of the descriptors (in technical terms, these structures behave as singularities). This means resulting models can take into account physical constraints, improving their accuracy. However, some of the most widely adopted descriptors also exhibit deformation patterns that do not change the descriptors but are not associated with a symmetry of the structure, these “spurious singularities” have a negative impact.

What inspired you to investigate this?

We have been fascinated by the geometric and mathematical problems associated with atomistic machine learning. This paper is, in a sense, a follow-up of a previous work of ours in which we observed a related type of behavior of descriptor. We believe that a better understanding of the mathematical underpinnings of these techniques can reveal some of the limitations of current frameworks and suggest approaches to overcome them.

What were the outcomes of your research?

We now have a better understanding of the link between molecular symmetries and the properties of representations, providing guiding principles to design better frameworks.

Why is open science important to you and your co-authors?

Machine learning is a fast-moving field and open research is essential to guarantee fast and inclusive research dissemination, meaning as many researchers as possible can quickly pick up our results and build on them.

What motivated you to publish with Open Research Europe?

We often complain about the limitations of the traditional publishing model, such as slow and opaque peer review and publishers who charge excessive fees to authors and/or libraries.

Open Research Europe seemed an interesting publishing venue, which we wanted to try and support. We were quite happy with the experience, and there will be more papers from us in the future.

What are your next steps?

This line of research is just one of many we are currently pursuing. Still, we believe there is more to discover about the relationships between geometry, the mathematical construction of descriptors, and the ultimate accuracy of machine learning models. We have a preprint showing similar pathologies affecting some types of geometric deep-learning models, suggesting that the problem is ubiquitous and extends beyond the class of representations we discuss here.

Want to find out more? Read the article here.