Physics-informed Generative Adversarial Networks Applied to Dichroic Filters' Properties Regression

Physica Scripta  ·  2025

Mateus Jesus de Oliveira Jesus de Oliveira Rodrigues; Maria Cecilia Queiroga Bazetto; Ana Amelia B. Machado; Mariama Rebello Sousa Dias; Gustavo do Amaral Valdiviesso; Anibal Thiago Bezerra;

HIGHLIGHT

While first-principles simulations are feasible,they are computationally costly and struggle to capture imperfections of real devices. To address this, we present the use of machine learning employing conditional physics informed generative adversarial network (CPIGAN) to simulate the optical transmittance of dichroic filters.

RELATED

Papers      Patents     Grants     Venues     Experts    Cite    Share    Save

ABSTRACT

Optical dichroic filters are key components of light detectors, and simulation is a desirable tool in detectors' research and development (R&D). Simulations must address the nuances of light-matter interactions to provide a realistic representation of the detector's performance. Accessing this information requires experimental data and precise measurements of the optical response under various conditions. While first-principles simulations are feasible,they are computationally costly and struggle to capture imperfections of real devices. To address this, we present the use of machine learning employing conditional physics informed generative adversarial network (CPIGAN) to simulate the optical transmittance of dichroic filters. Training on measured and simulated data enables the prediction of the filter's behavior, capturing deviations between experiment and simulation, thereby shifting the bulk of the computational effort to training. Through training on datasets with varying numbers of experimental curves, we demonstrated that employing an optimized CPIGAN architecture with physics-informed loss terms enables the generation of realistic synthetic transmittance curves, even in scenarios with limited experimental data. We demonstrated good training results by a trade-off between properly selecting the experimental representative samples and a robust hyperparameters' optimization. We foresee that the plasticity of the CPIGAN model will be essential for handling diverse filter designs, enabling reliable, effective, and precise ML-assisted simulations.

 Paraphrase     Questionfy     Wikify     Simplify     Summarize