Abstract
This paper presents a probabilistic surrogate model based on Conditional Normalizing Flows (CNFs) for the rapid prediction of radiative properties in nanoparticle-embedded layers. Traditional Monte Carlo (MC) simulations, while accurate, are computationally prohibitive for iterative design and inverse problems. The proposed CNF model, utilizing a RealNVP framework, provides high-fidelity spectral predictions of reflectance, absorbance, and transmittance while simultaneously offering principled uncertainty quantification through 95% confidence intervals. Results demonstrate excellent agreement with ground-truth MC simulations across a broad spectral range (0.2 to 2.5 μm), significantly reducing computational overhead.
Introduction
The interaction of light with nanoparticle-embedded materials is fundamental to applications in solar energy, thermal management, and biomedical imaging. Predicting these interactions typically relies on Monte Carlo (MC) simulations, which track millions of photon trajectories to resolve radiative transfer. However, the high computational cost of MC simulations limits their utility in high-throughput design optimization and real-time inverse modeling. While deterministic neural networks offer speed, they lack the ability to quantify prediction uncertainty—a critical requirement for robust engineering. This work introduces a Conditional Normalizing Flow (CNF) surrogate that addresses these limitations by providing both fast inference and a full posterior distribution of the optical outputs.
Methods
The researchers developed a surrogate model using Conditional Normalizing Flows (CNFs). Unlike standard regression models that provide point estimates, CNFs learn to map a simple base distribution (e.g., a standard Gaussian) to a complex target distribution through a sequence of invertible, differentiable transformations.
By conditioning these transformations on physical parameters—wavelength ($\lambda$), absorption ($\mu_a$), scattering ($\mu_s$), anisotropy ($g$), and particle size distribution ($\rho$)—the model captures the full posterior distribution $p(\text{optical output} \mid \text{physical inputs})$. The architecture employs a RealNVP (Real-valued Non-Volume Preserving) framework where a fully connected neural network (FCNN) serves as a conditioner to predict the shift and scale parameters required for the flow transformations.
Datasets
Training data was generated using the MCML (Monte Carlo for Multi-Layered Media) algorithm integrated with Mie theory to calculate the optical properties of individual nanoparticles. The resulting dataset covers a broad spectral range from 0.2 to 2.5 μm. The target outputs include full spectral profiles for total reflectance, absorbance, and transmittance. The model's performance was validated using 5-fold cross-validation to ensure robust generalization to unseen physical configurations.
Results and Discussion
The CNF surrogate demonstrates high accuracy and reliable uncertainty calibration. As shown in Figure 1, the mean spectral predictions for total reflectance, absorbance, and transmittance align almost perfectly with the ground-truth MC simulations. Notably, the model effectively captures sharp spectral transitions, such as the one occurring around 1.1 μm, suggesting it is expressive enough to handle the non-linearities inherent in radiative transfer.
Beyond point estimates, the model provides a 95% confidence interval (CI) for its predictions. Figure 2 illustrates these posterior distributions, where the true values consistently fall within the shaded CI regions. This capability allows researchers to identify spectral regions where the model (or the underlying physics) is less certain. However, the authors note a physical discrepancy: because the model is purely probabilistic, its confidence intervals can occasionally extend into negative values, which is physically impossible for reflectance and absorbance.
Conclusion
The CNF-based surrogate provides a powerful alternative to traditional MC simulations, offering both high-speed inference and principled uncertainty quantification. Once trained, the model can generate thousands of spectral predictions in a fraction of the time required for a single MC simulation. Future work will focus on embedding physical constraints, such as non-negativity and energy conservation (reflectance + absorbance + transmittance = 1), directly into the network architecture to improve physical consistency and extending the framework to more complex multi-layer systems.
Summary updated