PI: Michael Wilhelm

This project evaluates the ability of advanced generative AI (GenAI) models to interpret infrared (IR) reflectance spectra—an analytical task that is traditionally challenging because of particle‑size‑dependent morphology effects, structural isomerism, and the nonlinear behavior of reflectance signals. Unlike absorption spectroscopy, reflectance spectra are strongly influenced by sample structure, making direct database matching impractical. To address this, the project leverages PNNL’s extensive databases of optical constants and the ray tracing physics‑based simulator to generate synthetic reflectance spectra under controlled conditions. These simulated spectra serve as inputs for evaluating several OpenAI models (e.g., GPT‑4.1, GPT‑5, o3‑Pro, and o4‑mini), both with basic and advanced prompt engineering strategies. The study assesses how well these GenAI models can do the following:

• Parse spectral data
• Identify key peaks and functional groups
• Infer molecular features
• Predict material identity, thickness, and other simulation parameters
• Benefit from fine‑tuning strategies

The ultimate goal is to develop a PNNL‑trained GenAI model capable of reliably assigning complex real‑world reflectance spectra to their constituent chemical components by combining physics‑based spectral simulation with machine‑learning interpretive capabilities.