Adaptive deep learning for the prediction of laser-induced colours on evolving printable substrates

About this position

Context:

Laser-induced colouring has become a key technology for producing visual security features on high-value documents and products. By locally and precisely tuning the optical properties of a substrate through carefully controlled laser parameters, it is possible to generate a wide range of colours in a contactless, flexible and cost-efficient manner, and even to multiplex several images that can be revealed independently under different observation conditions (e.g. transmission versus reflection). Scaling up the technology requires to guarantee perfect reproducibility over time.

However, colour produced by a given set of laser parameters depends strongly on the initial optical state of the printable substrate. There is currently no physical model able to predict, quickly and reliably, the colours obtainable on a given sample, and rigorous electromagnetic simulations remain far too costly to be generalised at the scale of millimetre-large, quasi-random metasurfaces.

Topics:

This 18-month post-doctoral position, funded by Toppan Security and carried out at Laboratoire Hubert Curien, aims to improve the reliability of the laser colouring process using deep learning. The work targets a printable-substrate-plus-laser-process technology recently brought to market by Toppan Security, for which batch-to-batch reproducibility is not perfectly guaranteed by the fabrication process. The overall objective is to develop adaptive neural networks able to predict the spectra and colours produced by laser on samples that may evolve slowly over time, and to do so, ultimately, without performing any new laser-marking tests on new samples — relying instead on the optical properties of the initial (unprocessed) sample.

The adaptive neural network will be trained on experimental spectra extracted from hyperspectral images of many samples on which databases will be produced using a fine sampling of laser processing parameters. The successful candidate will build on preliminary results obtained within the LabHC teams Functional Materials and Surfaces and Machine Learning. Different transfer-learning strategies for deep learning (fine-tuning, representation learning, meta-learning, hypernetwork-based conditioning, spectral priors, etc.) will be explored in order to identify the most relevant solutions for this industrial use case.

Objectives:

• Developing and comparing adaptive deep-learning models that predict laser-induced colours on new substrates while accounting for their initial optical response, with the ultimate goal of a zero-shot adaptation requiring no new laser-marking tests.
• Measuring hyperspectral data on industrial samples.
• Exploiting hyperspectral characterisation of both the initial samples and the laser-produced spectra and colours as physical priors informing the prediction models.
• Defining the experimental protocol needed to guarantee reliable predictions — in particular the number of distinct samples required for training, the extent of sample variability that must be covered, and the number of colours to be laser-produced on each training sample.
• Assessing colour-prediction accuracy against perceptual metrics and demonstrating improved process reliability on the Toppan Security technology.
• The candidate will be invited to participate in the laser-induced printing of the databases by using a user friendly industrial device.
• In a further step of the work plan, the candidate will be involved in the development of optimization strategies to rapidly find the largest possible gamut of colors that can be laser-printed on various kinds of supports.