The rapid discovery of high-performance photocatalysts remains a critical challenge in sustainable chemistry, where traditional screening workflows are limited by high energy consumption and long experimental times. In this work, I introduce LuminaFemto AI, a simulation-based active-learning framework designed to autonomously identify efficient photocatalysts under an energy budget at the femtojoule scale. I generate synthetic spectral fingerprints that mimic UV–Vis absorption and photoluminescence profiles of candidate materials. A Gaussian-process surrogate model is trained to learn the nonlinear mapping between spectra and photocatalytic performance. At each iteration, an uncertainty-driven acquisition function selects the next most informative candidates to minimize both prediction error and cumulative energy cost. Numerical experiments on a library of 200 virtual materials show that LuminaFemto AI converges toward the optimal catalyst in fewer than 25 iterations, achieving a sub-1 h simulated discovery time while reducing per-experiment energy consumption by three orders of magnitude compared to random exploration. This framework establishes a quantitative connection between spectral information, learning efficiency, and energy-aware optimization, paving the way for autonomous, ultra-low-power laboratories for materials discovery. “Machine learning is transforming the field of materials discovery by enabling algorithms to learn to see, learn to estimate, and learn to search in compositional spaces previously inaccessible to human trial‑and‑error.” Keywords Active Learning Photocatalyst Discovery Femtojoule Optimization Spectral Analysis Gaussian Process Regression Autonomous Materials Discovery Energy-Aware Simulation Machine Learning in Chemistry High-Throughput Screening LuminaFemto AI