Predictive Generation of Microstructured Surfaces with Controlled Optical Properties Using Generative Adversarial Networks

This PhD position focuses on creating a predictive framework to generate microstructured surfaces with controlled optical properties using generative adversarial networks (GANs). The project aims to bridge the gap between microtopography and appearance (BRDF/NDF) without relying on iterative prototyping and measurement cycles. The research will explore GANs and neural networks to synthesize realistic height maps from user-defined parameters or synthetic data, ensuring alignment between virtual appearance and real-world manufacturability. Conducted within the ANR-funded APPEARANCE_ON_DEMAND project, this work will leverage collaborations between XLIM and iLM institutes, combining physical modeling, numerical optimization, and AI to establish a predictive design framework for surfaces with controlled visual appearance.

• Develop and train GANs to generate realistic height maps of microstructured surfaces from synthetic or experimental data.
• Integrate virtual measurements to guide surface generation, ensuring manufacturability and alignment with target BRDF properties.
• Design progressive algorithms for adaptive mesh densification and refinement while preserving optical and mechanical properties.
• Optimize computational efficiency through network pruning, model compression, and transfer learning.
• Conduct real-surface metrology using stereophotometry or confocal microscopy to validate generated surfaces.
• Collaborate with industrial partners to ensure generated surfaces meet fabrication constraints via Mihaly relief printing technology.
• Publish findings in peer-reviewed journals and present results at international conferences.