We refer to matter as “active” when the constituent particles convert internal energy into self-propulsion. Examples include biological systems like bacterial colonies and cell tissues, but also macroscopic systems like human crowds. Collections of active particles exhibit radically different behavior compared to their passive counterparts. A prime example is motility-induced phase separation (MIPS), where purely repulsive particles spontaneously condense into dense clusters solely due to self-propulsion.
Simulations provide a powerful tool to study the aforementioned non-equilibrium phenomena in a controlled manner. However, when probing different regimes of activity for the same system, standard physics-based simulations must be re-run for every parameter choice, which is computationally expensive. This severely limits our ability to map full phase diagrams and discover novel emergent properties across varying activity levels. Recently, machine learning approaches known as “thermodynamic maps” (https://doi.org/10.1073/pnas.2321971121) have been introduced to interpolate between ensembles based on fluctuation levels—for example, generating particle dynamics at intermediate temperatures using only data from high and low temperature samples. Similarly, dynamics under different chemical potentials can be generated.
In this project, you will explore how machine learning models can be adapted to map dynamics of active particles from one level of activity to another. You will both run physics-based simulations and train data-based diffusion models. The project is under the co-supervision of Sara Jabbari-Farouji, expert in active matter. Experience with particle-based simulations and/or machine learning is desirable.
