PHOTOM - Research project

 

 

The success of Machine Learning relies on the availability of a large amount of training examples. While most of vision and natural language processing tasks can benefit from massive data coming from the Internet, many real-world applications still require to resort to physics-based models to be addressed. This is particularly the case when predicting complex dynamical phenomena (e.g. climate forecasting, self-organization of matter, fluid dynamics, evolution of biological structures) which are typically modelled by partial/ordinary differential equations (PDE/ODE). Considering that most differential equations are costly over-simplistic representation of the underlying real phenomenon, a recent trend emerged aiming at combining physical knowledge and data-driven approaches. This combination can be addressed in different ways: (i) Machine Learning can be used to train surrogate models of PDE solvers. These latter are usually expensive (they can take weeks!) and therefore do not scale well in high dimensions. Therefore, learning a cheap approximation of PDEs might be of great interest in many real-world applications; (ii) Machine Learning can be used to discover automatically the dynamics (i.e. the underlying PDE) that is hidden behind the data. This can be done by pre-defining or learning the differential terms of PDEs and optimizing the parameters of a linear combination; (iii) More challenging, the physical knowledge can be integrated in some way during the learning process.

The objective of this thesis is to develop new methodological contributions in physics-guided Machine Learning in the specific domain of laser-matter interaction. Although this field has received some attention during the past few years, it is still an emerging and exciting topic. The optimization of inhomogeneous energy absorption of ultrafast laser light impinging a rough surface remains an ongoing issue for laser manufacturing. 3D topography, defined by shape, size and concentration of surface nano-reliefs, determines radiative and nonradiative contributions in light scattering and surface wave excitation. This can be modelled by computational electrodynamics through Finite Difference Time Domain approach to solve Maxwell's equations (coupled PDE). Machine learning will be devoted to anticipate extreme local field enhancement and collective effects that appear during light-surface coupling, considering adequate energy and flux conservation laws.