AI Driven Laser Parameter Search: Inverse Design of Photonic Surfaces using Greedy Surrogate-based Optimization

TL;DR

The paper tackles efficient inverse design of photonic surfaces to match target spectral emissivity with a limited number of experimental evaluations. It introduces ALPS, a greedy surrogate-based optimization that uses a Random Forest forward model with RF-PCA to predict emissivity from femtosecond-laser parameters and iteratively minimize RMSE to a target. Key contributions include a reusable forward surrogate enabling warm starting, batch greedy sampling to reduce evaluations, and demonstrated gains via cross-target and cross-material warm starting across synthetic and real photonic benchmarks. Empirical results show ALPS consistently outperforms a range of baselines (PSO, DE, MADS, NM, LBFGSB, BO) and achieves rapid, robust convergence, enabling resource-efficient autonomous photonic surface design.

Abstract

Photonic surfaces designed with specific optical characteristics are becoming increasingly important for use in in various energy harvesting and storage systems. , In this study, we develop a surrogate-based optimization approach for designing such surfaces. The surrogate-based optimization framework employs the Random Forest algorithm and uses a greedy, prediction-based exploration strategy to identify the laser fabrication parameters that minimize the discrepancy relative to a user-defined target optical characteristics. We demonstrate the approach on two synthetic benchmarks and two specific cases of photonic surface inverse design targets. It exhibits superior performance when compared to other optimization algorithms across all benchmarks. Additionally, we demonstrate a technique of inverse design warm starting for changed target optical characteristics which enhances the performance of the introduced approach.

Figures (22)

• Figure 1: Photonic surfaces inverse design segments and examples: (a) The ML experimental model pipeline developed to assess laser fabrication parameters. Each model is trained using distinct datasets and categorized according to the surface material employed for laser texturing. The RF algorithm predicts the PCA components, which are then transformed into spectral emissivity curves. (b) The inverse design loop comprises three phases: (1) Generation of laser parameters using the ALPS framework and assessment via an experimental model (defined in (a)) to produce a spectral emissivity curve (2), followed by (3) comparison of this curve against the target spectral emissivity and finally updating ALPS with new data for iterative decision making. (c) Example of the discrepancy, $\epsilon$, as defined in Eq. (\ref{['eqn:inverse_optimization']}), between the user-defined target spectral emissivity curve $\mathbf{y}$ and the spectral emissivity curve derived from evaluating the design vector $\mathbf{x}$ with the model defined in (a). (d) Photonic surfaces inverse design benchmark targets: TPV emitter (top) and the near-perfect emitter (bottom). The TPV emitter target switches to 0 emissivity at a wavelength of 4.6 $\upmu$m. The plain spectral emissivity curves are measured from the material without laser texturing.
• Figure 2: The inverse design results for the logistic growth (top) and sinusoidal oscillation with damping (bottom) benchmarks: (a) Convergence graphs for all optimization algorithms for the logistic growth benchmark. (b) ALPS solution reconstruction graph for the logistic growth benchmark. (c) Convergence graphs for all optimization algorithms for the sinusoidal oscillation with damping benchmark. (d) ALPS solution reconstruction graph for the sinusoidal oscillation with damping benchmark.
• Figure 3: The inverse design results for the Inconel photonic surface benchmarks. Convergence graphs for all optimization algorithms are shown in the first column, while in the second, ALPS solution reconstruction graphs are shown: (a) Convergence graphs for the Inconel near-perfect emitter target benchmark. (b) ALPS solution reconstruction graph for the Inconel near-perfect emitter benchmark. (c) Convergence graphs for the Inconel TPV emitter target benchmark. (d) ALPS solution reconstruction graph for the Inconel TPV emitter.
• Figure 4: The inverse design results for the Stainless steel photonic surface benchmarks. Convergence graphs for all optimization algorithms are shown in the first column, while in the second, ALPS solution reconstruction graphs are shown: (a) Convergence graphs for the Stainless steel near-perfect emitter target benchmark. (b) ALPS solution reconstruction graph for the Stainless steel near-perfect emitter benchmark. (c) Convergence graphs for the Stainless steel TPV emitter target benchmark. (d) ALPS solution reconstruction graph for the Stainless steel TPV emitter.
• Figure 5: The inverse design results of ALPS with and without cross-target warm starting for the photonic surface benchmarks (ALPS$_{ws=True}$ means ALPS with warm starting and is shown as the red line). Convergence graphs are shown in the first column, while in the second, the solution reconstruction graphs are shown: (a) Convergence graphs for the Inconel near-perfect emitter target benchmark. (b) Solution reconstruction graph for the Inconel near-perfect emitter benchmark. (c) Convergence graphs for the Inconel TPV emitter target benchmark. (d) Solution reconstruction graph for the Inconel TPV emitter.