Machine learning: The key to faster and easier calculations of free energy

What is it about?

The free energy of a system is a measure of its capacity to do work. Calculating the free energy of a system is important in the study of natural sciences. A commonly used math strategy for this is called free energy perturbation (FEP). But FEP can only work if there is sufficient overlap between the distributions of the thermodynamic states of the system. Targeted FEP (TFEP) is an "invertible mapping" approach to increasing overlap. This essentially means that for two thermodynamic states A and B, TFEP transports points from A to a new distribution A', which hopefully overlaps with B. Crafting this mapping is a huge challenge for a person, but not for machine learning (ML) techniques. This study turns TFEP into an ML problem. It describes a deep neural network that is optimized to maximize the overlap.

Featured Image

Photo by Kevin Ku on Unsplash

Photo by Kevin Ku on Unsplash

Why is it important?

Calculating free energy is key to unraveling the properties of nanoscale systems. It saves time during research and development. But it is hard to accurately estimate these free energies. They cannot be directly measured or estimated indirectly from other measurements. Although TFEP is a powerful tool to calculate free energies, mapping the system and increasing overlap is hugely complex. This limits its use in research. Using ML to perform the mapping makes the application of TFEP simpler and faster. This ultimately makes research faster. This neural network also respects "symmetries" in the system structure. This makes it useful for estimating the free energy of nanoscale systems. Accounting for these symmetries is a big challenge in ML problems. The ML technique also provides more precise estimates of free energy, without needing any additional data to train the neural network. KEY TAKEAWAY: TFEP is a powerful tool for calculating the free energy of a system. But its uptake is limited by the enormous complexity of the mapping. ML techniques can be used to maximize overlap and make mapping simpler and faster.