Statistical Mechanics Approaches to Digital Twins

A digital twin is a dynamic, computational model that serves as a live, evolving counterpart to a physical system. At the heart of a digital twin is a high-fidelity simulation of that system. Reliably constructing such simulations presents challenges that are both technical and fundamental. I will discuss recent efforts to automate the construction of high-fidelity physical models underlying digital twins in diverse application domains, along with the associated challenges. In particular, as models grow in scope and fidelity, they inevitably produce a proliferation of unknown parameters. Statistical mechanics approaches to models with many parameters have documented a universal phenomenon of information compression, sometimes known as sloppiness, places fundamental limits on parameter identifiability in complex systems. The talk will highlight how this landscape of parameter uncertainty complicates inference at scale, and how information-theoretic methods can be used to quantify and manage these limitations. I will then discuss strategies for parameter reduction that align model complexity with the information content of data, thereby connecting microscopic descriptions to predictions of interest. Finally, I will outline how these approaches enable the construction of digital twins that are both tractable and predictive, illustrating how insights from statistical mechanics can guide the design of high-fidelity, information-efficient simulations across the physical sciences.