Assessing ANI neural network potentials for geometry optimization of organic macrocycles

کد مقاله : 1153-ICOC

نویسندگان

محمد حسین حورزاد *1، امیر بلوچ2، علی رضا باوفا طرقبه3 1دانشگاه تربیت مدرس 2دانشگاه علوم تحقیقات 3دانشگاه صنعتی شریف

چکیده مقاله

The computational prediction of equilibrium molecular geometries plays a pivotal role in the rational design of functional organic materials. While high-accuracy quantum mechanical methods such as density functional theory (DFT) are reliable, their steep computational cost limits applicability to large or conformationally flexible systems like organic macrocycles [1]. Recent advances in machine learning have introduced neural network interatomic potentials—particularly the ANI (Accurate Neuronal Network Interaction) family—as efficient surrogates that approach DFT accuracy at a fraction of the cost [2]. In this work, we systematically evaluate the performance of four official ANI variants—ANI-1, ANI-1x, ANI-1ccx, and ANI-2x—for geometry optimization of H/C/N/O-containing macrocycles. Reference structures were generated using DFT (ωB97X-D/def2-SVP). We tested multiple optimization algorithms, including L-BFGS, FIRE, BFGS, and conjugate gradient, while varying convergence criteria (force tolerance: 10⁻²–10⁻⁴ eV/Å), maximum iterations (100–2000), and, where applicable, MD-based time steps (0.5–1.0 fs). Notably, time step is only relevant for dynamics-based optimizers like FIRE and has no role in pure minimizers such as L-BFGS.

کلیدواژه ها

ANI potentials, macrocycles, geometry optimization, machine learning, quantum chemistry

وضعیت: پذیرفته شده