Scientific direction Development of key enabling technologies
Transfer of knowledge to industry

PhD : selection by topics

Technological challenges >> Numerical simulation
3 proposition(s).

See all positions [+]

Dynamic Mesh Reordering by Cache-aware Heuristics for scientific computing

Département Systèmes et Circuits Intégrés Numériques

Laboratoire pour la Confiance des sYstèmes de calcuL



Numerical simulation (.pdf)

In the context of simulation codes on unstructured grids, two elements appears: the traversal of the mesh on a CPU appear as random, and the "unstructured" aspect of the mesh makes accelerators problematic. However, such meshes allow the simulation at multiple scales, replacing multiple nested regular grids by a single structure, for example in tsunami simulations, resulting in a more compact structure and decreasing the computational power needed. Additionally, in some contexts the mesh will evolve in time, making partionning for multiple compute units difficult with their associated caches. This thesis focuses on the development of cache-aware heuristics of traversal of unstructured meshes, so as to allow on-the-fly reordering, in particular when sending compute kernels and mesh data on accelerators, so as to start computing before the mesh transfert has ended. This thesis will rely on existing results on mesh partitionning with tools such as SCOTCH and Metis, on models of memory / cache hierarchies and transfert capabilities (accelerators), and space filling curves related work (Hilbert-Peano, Sierpinski). The focus will be on the TsunAWI and FESOM (Finite Element/volume Sea-Ocean Model) codes from AWI and OpenFOAM. The student may have the opportunity to spend time at AWI (Germany) during the PhD.

Download the offer (.zip)

Development of innovative algorithms for creating non-intuitive masks using neural networks for grayscale lithography

Département des Plateformes Technologiques (LETI)




Numerical simulation (.pdf)

The realization of micronic 3D structures makes it possible to manufacture key functional elements of microelectronics such as micro-lenses for optical imagers. These lenses can in particular be produced using a resin flow process or by gray-scale lithography (grayscale). Grayscale lithography offers the advantage of being able to create structures of different topographies in a single process step. Its success strongly depends on the correctness of the modeling of the process and of the optimization strategy of the optical mask. Grayscale lithography has been developed at CEA-LETI over the past 3 years as part of an industrial collaboration [1]. These developments have made it possible to produce state-of-the-art results [2]. CEA-LETI wishes to continue this research work towards new design and data preparation methodologies. Artificial intelligence and neural networks in particular open up a wide range of possibilities in this area. A first promising study has been carried out in this direction, and shows the full potential that such a technique can offer for the creation of masks, if we break free from classical algorithms. The emergence of multi-beam electronic writing tools for the manufacture of optical lithographic masks makes it possible now to consider the use of curved shapes. It is therefore possible to think about changing the algorithms to use non-regular shapes on the mask for the search for ideal optimal solutions. Necessary during the learning stage, the modeling of the lithography process will also be an important part of the thesis and will follow on from previous work [1]. [1] Thesis of P. Chevalier, Study of a 3D micro-fabrication method for microlenses imaging application (2021) [2] P. Chevalier et al., Rigorous Model-Based Mask Data Preparation, IEEE JMEMS (2021) Applications have to be sent to: Sébastien Bérard-Bergery : Loic Perraud :

Download the offer (.zip)

Qualitative reasonning and design of complex systems

Département Ingénierie Logiciels et Systèmes (LIST)

Labo. ingénierie des langages exécutables et optimisation



Numerical simulation (.pdf)

The design of complex systems is an activity that affects many industrial and research fields. This implies difficulties in modeling and simulating by the heterogeneous nature of the data involved, with discrete and continuous aspects. Two approaches are possible. Quantitative methods, whose analyzes are numerical, are the most used: their results are precise but they consume a lot of time and resources. Qualitative methods are based on a symbolic interpretation of the models, and can be used without knowing all the numerical parameters, by relying on dependency relationships between variables. They are less precise but they can be applied very early in the design phase and can be used to plan numerical simulations according to the objectives and to improve the results of analyzes (proofs, optimization, etc.). The work already carried out at the LIDEO laboratory of CEA LIST on modeling and qualitative simulation will be extended by the integration of concepts from naive physics and common sense reasoning to lead to an approach closer to engineering concepts. The results will be used for modeling, simulation but also for optimization on case studies representative of industrial examples.

Download the offer (.zip)

See all positions