Machine Learning (ML) is quickly providing new powerful tools for physicists and chemists to extract essential information from large amounts of data, either from experiments or simulations. Significant steps forward in every branch of the physical sciences could be made by embracing, developing and applying the methods of machine learning to interrogate high-dimensional complex data in a way that has not been possible before.
As yet, most applications of machine learning to physical sciences have been limited to the “low-hanging fruits,” as they have mostly been focused on fitting pre-existing physical models to data and on discovering strong signals. We believe that machine learning also provides an exciting opportunity to learn the models themselves–that is, to learn the physical principles and structures underlying the data–and that with more realistic constraints, machine learning will also be able to generate and design complex and novel physical structures and objects. Finally, physicists would not just like to fit their data, but rather obtain models that are physically understandable; e.g., by maintaining relations of the predictions to the microscopic physical quantities used as an input, and by respecting physically meaningful constraints, such as conservation laws or symmetry relations.
The exchange between fields can go in both directions. Since its beginning, machine learning has been inspired by methods from statistical physics. Many modern machine learning tools, such as variational inference and maximum entropy, are refinements of techniques invented by physicists. Physics, information theory and statistics are intimately related in their goal to extract valid information from noisy data, and we want to push the cross-pollination further in the specific context of discovering physical principles from data.