ETD

• carotenoids
• conjugated systems
• gaussian process regression
• lutein
• molecular dynamics
• neural network potentials

An accurate quantum mechanical description of molecular dynamics is essential for understanding photoactive systems such as carotenoids, whose extended conjugated backbone gives rise to complex electronic structures with pronounced multireference character. As a consequence, reliable sampling of their potential energy surfaces becomes computationally demanding with conventional electronic structure methods. In this work we investigate whether machine learning models can reproduce the ground state potential energy surface of lutein in the gas phase as a prototypical carotenoid. We compare Gaussian process regression (GPR) and neural network potentials (NNPs), analyzing their accuracy, scalability and robustness. While GPR is limited by memory scaling and NNPs require large representative datasets to ensure stability, we show that these challenges can be mitigated through optimized dataset construction and active learning strategies. The resulting models enable stable molecular dynamics simulations and reproduce structural, vibrational and optical properties in good agreement with quantum mechanical references, at a substantially reduced computational cost.