PD-TNN: Physics-Driven Tandem Neural Network for Color Prediction in Laser-Induced Thin Films
Yun-Jie Jhang, Tsung‐Ming Tai, Chia-Hung Chou, Wisarud Yongbanjerd, Pin-Han Huang, Hung‐Wen Chen · ACS Photonics · 2025
High Resolution Image Download MS PowerPoint Slide Accurately predicting laser machining parameters for thin-film coloration remains challenging due to the nonlinear and one-to-many mapping between parameters and resulting colors. Traditional methods rely on extensive trial-and-error experimentation, making them inefficient for practical applications. Moreover, existing deep learning approaches often neglect critical physical information and fail to align the optimization process with human color perception. In this work, we present a physics-driven tandem neural network (PD-TNN) that integrates the physics information and perceptual optimization to improve color prediction. By incorporating physics-based features–average fluence F and average time τ-PD-TNN mitigates the one-to-many mapping issue inherent in the inverse problem, leading to more physically consistent predictions. Additionally, employing Δ E as the loss directly aligns the optimization objective with human perception, enhancing accuracy beyond conventional numerical loss functions. PD-TNN achieves a mean test Δ E of 1.48, with 74% of predictions perceptually indistinguishable from targets (Δ E ≤ 2.3) and 97% within industrial standards (Δ E ≤ 7). Experimental results further confirm its ability to translate digital color values into laser-induced coloration on titanium surfaces. By overcoming the inefficiencies of both trial-and-error methods and conventional deep learning models, PD-TNN provides an effective approach for precision manufacturing and artistic applications, demonstrating its potential in laser technologies.