[Beautiful story] INRIA MALICE project: AI to unravel the mysteries of laser-matter interaction

Artificial intelligence is regularly in the news, thanks to the many possibilities it offers, particularly through machine learning. Its fields of application seem infinite, from healthcare to automotive, security and industry. And in recent years, a new field of research in this field has emerged: physics-informed machine learning, i.e., automatic learning that no longer relies solely on data, but also on physical knowledge, modeled by equations. Programs developed in this way can, for example, predict the behavior of a liquid in accordance with the laws of fluid mechanics.

Machine learning and surface engineering

" Current work of this type generally focuses on well-modeled physical phenomena, with the main aim of speeding up the resolution of equations," notes Marc Sebban, Professor of Computer Science at Université Jean Monnet Saint-Étienne and head of the MALICE project team. " This is the case, for example, in fluid mechanics and heat transfer. But machine learning could also prove extremely useful for learning more about mechanisms that we can only partially model.

This is the aim of the Inria MALICE(MAchine Learning with Integration of surfaCe Engineering knowledge) project-team, housed at the Hubert Curien laboratory, UMR CNRS, to which Télécom Saint-Étienne, a member of the Carnot Télécom & Société numérique institute, belongs. Officially created in December 2023, it aims to study the mutual contributions of artificial intelligence and surface engineering. It brings together experts in computer science and applied mathematics: six professors from Université Jean Monnet Saint-Étienne and two full-time researchers from Inria, along with a dozen PhD and post-doctoral students.

And to provide physics knowledge, the team can draw on a special feature of the Hubert Curien laboratory. " Since its creation in 2006, our structure has brought together digital specialists and physicists in the same place," points out Marc Sebban. " This rare specificity gives us a considerable advantage when it comes to carrying out our research work. We have over fifty physicists working in the offices next door to ours!

The puzzle of laser-matter interaction

This collaboration between the worlds of machine learning and surface engineering focuses on one physical phenomenon in particular: laser-matter interaction. " When you send an ultra-short laser beam onto the surface of a material, multiple nanoscale modifications take place," explains Marc Sebban. " Among the phenomena the laboratory studies is that of self-organization: by repeating laser pulses, matter organizes itself to make patterns appear from the many local interactions. "

While this phenomenon can easily be observed through a microscope, its mechanisms remain poorly understood. And with good reason: they involve a combination of physical phenomena relating to thermodynamics, fluid mechanics and wave propagation. So complex is this behavior that it is currently only incompletely modeled, through an equation that only partially explains the problem. All the more so as the variety of patterns generated is virtually infinite: they depend not only on the material, but also on the laser parameters - its energy, the number of pulses, the delay between each one... And the slightest change in a single property is enough to produce a totally different result.

The challenge of data availability

The MALICE project-team aims to gain a better understanding of this self-organizing phenomenon of laser-matter interaction, using machine learning. But beyond partial knowledge, this problem is accompanied by a major pitfall for machine learning: the lack of data. Data is the essential fuel for a machine learning model during its training phase - the stage during which it "learns" to perform the requested task, before being able to produce results on its own. However, for such a physical mechanism, producing large quantities of data would mean repeating multiple experiments over and over again, which would be time-consuming, costly and technically complex.

" That's why we intend to develop hybrid machine learning models , capable of relying on both a partial level of physical knowledge and a low volume of data," announces Marc Sebban. " To do this, we'll be using new artificial neural network architectures, whose behavior we'll need to analyze not only from an algorithmic point of view, but also from a theoretical one. " The idea is thus to compensate for the lack of data by the contribution of modeling equations - even imperfect ones - and, conversely, to compensate for partial knowledge of the physical phenomenon by the presence of data - even in small quantities.

In order to produce more data, the MALICE project-team has already begun working with the laboratory's physicists to set up new experimental protocols. They will also be using " data augmentation ", a technique involving the generation of new data from that already available, for example by adding noise. " Nevertheless, we will never have enough data to train our algorithms with a completely data-driven approach," moderates Marc Sebban. " Hence the interest in developing hybrid machine learning models , which also rely on physical equations to ensure the consistency of predictions. "

A win-win collaboration model

The issues tackled by the MALICE project-team are therefore not lacking in challenges. But for its leader, it is this complexity that makes the collaboration mutually enriching. " We're not here to serve the physicists: they bring us value too," explains Marc Sebban. " In this case, the pitfalls raised by laser-matter interaction are leading us to develop new methodological contributions in artificial intelligence, via our hybrid machine learning models . And, conversely, we are of course helping physicists to better understand the physical phenomena involved. "

Indeed, the research team's work will help to deepen current knowledge, by completing existing equations and even discovering new ones. The researchers will also be looking at transfer learning, which involves answering the following question: is the knowledge already acquired about one material transferable to another, without having to start from scratch? This is a common problem in machine learning, and one in which the team has already developed expertise.

Possible applications in many fields

MALICE also aims to tackle the problem of laser-matter interaction from a different angle. " Instead of predicting how matter will self-organize when it comes into contact with a laser, why not consider solving the opposite problem? " Starting with a pattern that the physicist would like to obtain, our models would then determine the laser parameters enabling it to be generated. "

This capability could prove invaluable in surface engineering. Indeed, the patterns formed by the self-organization of matter can confer new properties on the material: aerodynamics, hydrophobicity, impermeability, etc. It is thus possible to envisage new types of surface. This could lead to applications in the automotive industry, with "nano-holes" to help lubricate mechanical parts, or in healthcare, with "nano-peaks" to reinforce the material's protection against viruses and bacteria. Similarly, knowledge of laser-matter interaction can be used to secure and control the authenticity of identity documents, by encoding information in them. The Hubert Curien laboratory is already working on this issue as part of a joint laboratory with HID Global, to which the MALICE project team is contributing with two theses.