In-situ 4D tracking of microstructural evolution in atomistic simulations
The exponential growth of high-performance computing has enabled atomistic simulations involving billions or even trillions of particles, offering unprecedented insight into complex physical phenomena. However, these simulations generate massive amounts of data, making storage and post-processing increasingly restrictive. To overcome this limitation, on-the-fly (in-situ) analysis has emerged as a key approach for reducing stored data by extracting and compressing relevant information during runtime without significantly affecting simulation performance.
In this context, tracking the four-dimensional (space and time) microstructural evolution of materials under extreme conditions is a major scientific challenge. Atomistic simulations provide a unique spatial resolution to analyze crystalline defects such as dislocations, twinning, vacancies and pores, which govern dynamic phase transformations, melting, damage and mechanical behavior. By capturing their spatio-temporal evolution, it becomes possible to study their statistics, correlations and collective effects in out-of-equilibrium regimes, leading to more accurate predictive material models.
This project builds on advances of the exaNBody high-performance computing platform and a recently developed in-situ clustering method in the ExaStamp molecular dynamics code at CEA. This method projects atomic information onto a 3D Eulerian grid to perform real-time clustering. The objective is to extend this approach to full 4D tracking, enabling the time-resolved monitoring of clusters. This will allow dynamic graph-based analysis of their evolution, including changes in size, shape and temporal behavior, providing new insights into microstructural dynamics at the atomic scale.
Spin-lattice interactions in Machine Learning assisted ab initio simulations
The scientific field addressed by this postdoctoral project lies at the intersection of ab initio molecular dynamics, machine learning, and the thermodynamic characterization of materials under extreme conditions. Traditional AIMD simulations are a powerful tool to study temperature- and pressure-dependent properties from first principles, but their high computational cost limits their widespread use. By developing and applying machine learning-assisted sampling techniques like MLACS, this postdoc aims to drastically reduce the computational burden while retaining ab initio accuracy. This enables the efficient exploration of phase diagrams in high-pressure and high-temperature conditions. This research contributes to both fundamental understanding and practical modeling of materials, offering high-impact tools for the scientific community.
Bayesian inference-based ab initio phase diagrams
The scientific field addressed by this postdoctoral project lies at the intersection of ab initio molecular dynamics, machine learning, and the thermodynamic characterization of materials under extreme conditions. Traditional AIMD simulations are a powerful tool to study temperature- and pressure-dependent properties from first principles, but their high computational cost limits their widespread use. By developing and applying machine learning-assisted sampling techniques like MLACS, this postdoc aims to drastically reduce the computational burden while retaining ab initio accuracy. This enables the efficient exploration of phase diagrams in high-pressure and high-temperature conditions. This research contributes to both fundamental understanding and practical modeling of materials, offering high-impact tools for the scientific community.
Method for dimensionality reduction applied to deformed coupled cluster theory
Ab initio calculations in nuclear physics have undergone considerable progress over the past 20 years, enabling the study of several hundred nuclei with approximately 5% precision, notably through the PAN@CEA collaboration (A-Nucleon Problem at CEA) between DAM, DRF, and DES. These methods connect nuclear phenomenology to QCD theory via chiral effective field theory (cEFT) and find applications in both nuclear structure and particle physics.
Despite these advances, the majority of the Segrè chart remains inaccessible, with limitations to nuclei of mass A~100. This limitation stems from the computational and memory costs that scale with the desired mass and precision, related to the storage of large tensors. Recent research has demonstrated that a significant portion of the information in these tensors can be compressed through dimensionality reduction methods without significant loss of precision.
The postdoctoral project aims to extend these methods to the non-perturbative framework of deformed coupled cluster theory (dCC). The objectives are: 1) to implement the dCCSD method for nuclei up to A~80, 2) to develop its factorized version (TF-dCCSD) and validate it, 3) to extend it either to excited states (EOM-dCCSD) or to sub-percent precision (dCCSDT).