Huang Chi-Yen, Zhang You-Lun, Liao Su-Yu, Huang Wen-Chun, Chen Jiann-Heng, Dong Bo-Chang, Hsu Che-Ju, Huang Chun-Ying
Graduate Institute of Photonics, National Changhua University of Education, Changhua 50007, Taiwan.
Department of Applied Materials and Optoelectronic Engineering, National Chi Nan University, Nantou 54561, Taiwan.
Nanomaterials (Basel). 2025 Aug 21;15(16):1289. doi: 10.3390/nano15161289.
Accurate determination of the liquid crystal (LC) cell gap after filling is essential for ensuring device performance in LC-based optical applications. However, the introduction of birefringent materials significantly distorts the transmission spectrum, complicating traditional optical analysis. In this work, we propose a lightweight machine learning framework using a shallow multilayer perceptron (MLP) to estimate the cell gap directly from the transmission spectrum of filled LC cells. The model was trained on experimentally acquired spectra with peak-to-peak interferometry-derived ground truth values. We systematically evaluated different optimization algorithms, activation functions, and hidden neuron configurations to identify an optimal model setting that balances prediction accuracy and computational simplicity. The best-performing model, using exponential activation with eight hidden units and BFGS optimization, achieved a correlation coefficient near 1 and an RMSE below 0.1 μm across multiple random seeds and training-test splits. The model was successfully deployed on a Raspberry Pi 4, demonstrating real-time inference with low latency, memory usage, and power consumption. These results validate the feasibility of portable, edge-based LC inspection systems for in situ diagnostics and quality control.
准确测定填充后液晶（LC）盒的盒厚对于确保基于液晶的光学应用中的器件性能至关重要。然而，双折射材料的引入会显著扭曲透射光谱，使传统光学分析变得复杂。在这项工作中，我们提出了一种轻量级的机器学习框架，使用浅层多层感知器（MLP）直接从填充LC盒的透射光谱中估计盒厚。该模型使用峰峰值干涉测量法得出的实验获取光谱的真实值进行训练。我们系统地评估了不同的优化算法、激活函数和隐藏神经元配置，以确定一个在预测准确性和计算简单性之间取得平衡的最优模型设置。性能最佳的模型使用具有八个隐藏单元的指数激活函数和BFGS优化，在多个随机种子和训练-测试分割中，相关系数接近1，均方根误差低于0.1μm。该模型已成功部署在树莓派4上，展示了具有低延迟、内存使用和功耗的实时推理。这些结果验证了用于现场诊断和质量控制的便携式、基于边缘的LC检测系统的可行性。
