Research
This page gives a glimpse of the physical phenomena I investigate and my approach to research.
Glass transition
The theoretical description of the glass transition is one of the main open problems in condensed matter physics. Why does the viscosity of supercooled liquids increase dramatically by decreasing temperature? Why are some liquids good glass-formers and do not crystallize when supercooled, while others not? What is the origin of the strong correlations that characterize the motions of atoms and molecules close to the glass transition? To answer these questions, several theories have been developed but none of them provides yet a fully satisfactory and predictive description.
My efforts are devoted to recompose the facets of the glass transition phenomenology into a coherent picture. I use computer simulations of simple and yet realistic models of glassy liquids to ask "crucial" questions to theories and see which ones give sensible answers.
- D. Coslovich, A. Ninarello, L. Berthier, SciPost Physics 7, 077 (2019) [arXiv]
- D. Coslovich, M. Ozawa, W. Kob, The European Physical Journal E 41, 62 (2018) [arXiv]
- L. Berthier, G. Biroli, D. Coslovich, W. Kob, C. Toninelli, Physical Review E 86, 031502 (2012) [arXiv]
Sampling methods and machine learning
Molecular dynamics and Monte Carlo simulations are essential tools to understand the microscopic nature of condensed matter. My research group focuses on classical simulation and optimization methods, and develops a Python framework to facilitate data analysis and the implementation of new algorithms. However, the time scales accessible to conventional simulations are too limited to study the dynamics of supercooled liquids close to the glass transition: even if we employ high performance computing (HPC) resources, simulating one second of the microscopic dynamics of a liquid would require centuries of wall clock time.
One of our goals is to develop efficient simulation methods and fill the colossal gap between the time scales accessible to simulations and experiments. For instance, using the swap Monte Carlo method we accelerated the simulation of glassy colloidal suspensions by several billions of times and measured their structure and thermodynamics under extreme supercooling conditions. We also develop leight-weight but realistic force-fields to simulate supercooled liquids and glasses. We analyze their structure, dynamics and glass-forming ability using statistical physics and machine learning methods. In the long term, our efforts aim at identifying the correct theoretical description of the glass transition.
- D. Coslovich, R. Jack, J. Paret, The Journal of Chemical Physics 157, 204503 (2022) [arXiv]
- J. Paret, R. Jack, D. Coslovich, The Journal of Chemical Physics 152, 144502 (2020) [arXiv]
- A. Ninarello, L. Berthier, D. Coslovich, Physical Review X 7, 021039 (2017) [arXiv]
Heterogeneity of glassy materials
The slow structural rearrangements that take place in highly viscous liquids are characterized by non trivial spatio-temporal fluctuations, known as dynamic heterogeneities. The typical time and length scales associated to these fluctuations have been the subject of intense research since the early 2000s, as they are crucial to understand glassy dynamics at a fundamental level. Despite several claims, however, there is currently no consensus on the correct physical description of these fluctuations and their temperature dependence.
A fraction of such dynamic fluctuations can be explained using structural information only. This is achieved by considering an iso-configurational ensemble of independent trajectories starting from a given configuration. The iso-configurational average of the particle displacements still displays non-trivial spatial fluctuations, illustrated below for a glassy Lennard-Jones mixture. Predicting from minimal hypothesis the time dependence of the iso-configurational mobility is a key open challenge.
- D. Coslovich, A. Ikeda, The Journal of Chemical Physics 156, 094503 (2022) [arXiv]
- G. Hocky, D. Coslovich, A. Ikeda, D. Reichman, Physical Review Letters 113, 157801 (2014) [arXiv]
Soft condensed matter
When the typical size of the elementary constituents becomes mesoscopic, materials become soft: when subject to deformation, they flow more easily than hard matter made of atoms and molecules. Soft materials, like mayonese, ink or foams, are omnipresent in our daily life and display a fascinating and complex physical behavior.
Colloidal suspensions are a prototypical example of soft condensed matter: they are composed by large macromolecules, whose sizes range from the nanometer to the micrometer scale, suspended in a microscopic solvent made of smaller molecules (ex. water). From a fundamental viewpoint, colloids are ideal model systems because of the possibility to tailor their mutual interactions, for instance by controlling their internal architecture or by altering the solvent properties. Understanding the link between these physical parameters and phase behavior or self-assembly is one of the most challenging goals of soft condensed matter physics.
My research on soft matter is based computational models that employ effective interactions. For a broad class of macromolecular systems, such as linear chains, dendrimers or microgels, the effective interactions are ultrasoft: they remain finite irrespective of the separation between the centers-of-mass of the macromolecules. Ultrasoft particles display a more complex phase behavior than hard ones as well as peculiar glassy dynamics, which can bring insights into fundamental questions about the glass transition.
- D. Coslovich, A. Ikeda, K. Miyazaki, Physical Review E 93, 042602 (2016) [arXiv]
- D. Coslovich, M. Bernabei, A. Moreno, The Journal of Chemical Physics 137, 184904 (2012) [arXiv]
- D. Coslovich, J. Hansen, G. Kahl, The Journal of Chemical Physics 134, 244514 (2011) [arXiv]
Reproducible research
Over time, I became increasingly concerned by how we do computational research in physics. What is needed to make a computational science project reproducible? Is it worth to make it fully reproducible? These issues are hard to address not only because of the technical hurdles on the way to reproducibility - such as workflow management or data provenance and persistency: reproducible research requires time and effort, while the dominant "publish or peril" academic paradigm pushes for high publication rates.
There are several excellent packages and platforms for reproducible computational research. Still, I am more inclined toward a DIY, low-level approach: some bare-bones templates for reproducible projects, based on Python scripts or notebooks, are available on my git repository. Packages and Docker images to fully reproduce the research findings of some recent papers of mine are available on Zenodo - an EU-funded, large-scale platform for data persistency.