Molecular Simulation for the Next Decade

Abstract

Molecular Simulation is nowadays one of the most powerful tools for investigation of matter. Its results have led to fundamental discoveries with the effect of forging current technology and shaping its future beyond any expectation. From molecular docking [1] to sophisticated materials, [2] simulation can pilot the design of futuristic systems for the most advanced laboratories. However its physical models and numerical techniques need to be continuously upgraded in order to have a constructive interplay with rapidly evolving experimental techniques. Removing barriers between different length and time scales was the desired target at the beginning of the new millennium. [3] The rationale behind this aim lies in the fact that molecular simulation is a quantitative tool that can satisfactorily determine both the microscopic origins of large scale properties and the influence of large scale behavior on the microscopic scale. Only knowledge of the interplay of scales can guide the accurate design of atomic or molecular systems with properties on demand. The program put forward in Ref. [3] has been implemented in many research centers all over the world, with relevant results in several fields of advanced research, from molecular biology, [4] to materials science, [5] to the design of novel "green" substances of high technological impact such as ionic liquids, [6] to cite a few. Current molecular simulation routinely links different scales in space and time. Moreover the idea that the governing physical principles of simulated models should be framed, as much as possible, in a rigorous mathematical language is starting to become a standard procedure for many researchers, with the consequence that simulations are characterized by an increasing physical accuracy and reproducibility. [7] Such conceptual progress is taking place in parallel to what can be considered a true technical revolution in the field, the contribution of artificial intelligence. Machine learning is not only increasing in a substantial manner the numerical efficiency of simulation algorithms but is also contributing to the refinement of atomic and molecular models where the currently available ones do not offer the desired accuracy. [8] However, it should be kept in mind that machine learning approaches, while being more efficient than other numerical approaches, still need well defined physical models. In addition, there is a key question regarding the generalization properties of artificial neural networks, i.e. when and why machine learning algorithms can (or cannot) generalize from the training data. In my view the major challenge for the future consists in finding an optimal balance between the research that develops first principle models and the research that empowers the technical efficiency of ma-