An international team of researchers has developed an artificial intelligence model that can predict the motion of atoms in molecular systems directly, dramatically speeding up simulations used in chemistry, physics, and materials science. The model, called MDtrajNet, bypasses the traditional step-by-step calculations of atomic positions, a process that can require millions of steps and substantial computing power.
“For years, researchers have been trying to describe the behaviour of matter at the atomic scale with increasing precision,” said Pavlo O. Drala, lead author and physicist at the Institute of Physics, Faculty of Physics, Astronomy, and Applied Computer Science, Nicolaus Copernicus University in Toruń. “When chemists want to understand how a molecule vibrates, how it changes shape, how a reaction runs, or how atoms are arranged in a crystal lattice, they turn to molecular dynamics simulations. This is a crucial but expensive method.”
Traditional molecular dynamics simulations require moving a system forward in very small increments and calculating forces and positions after each step. MDtrajNet instead predicts the positions of atoms for a given future moment, bypassing the intermediate steps and rethinking the computational logic.
The model combines two advanced AI approaches: it incorporates spatial symmetry to recognize that molecules remain unchanged when rotated or shifted, and it uses transformer architectures, typically employed in language models, to analyse relationships between atoms. It takes as input atom types, initial positions, velocities, and prediction time, and outputs the system’s future configuration.
The research team built a baseline model, MDtrajNet-1, and trained it on 173 molecular systems containing two to nine atoms. “We achieved a nearly hundred-fold speedup compared to classic machine learning simulations trained on the same data, without sacrificing any accuracy,” the study notes. The model produced trajectories consistent with reference data and remained stable for molecules not seen during training.
Tests included the alanine dipeptide, a standard system in computational chemistry and biophysics. The AI accurately reproduced key conformational regions, which could have implications for drug design, materials development, catalysts, and batteries.
The researchers caution that the model is still limited, covering only small systems and a portion of chemical space. Some trajectories were unsuccessful for underrepresented cases in the training set. Nevertheless, the results show that molecular dynamics simulations can be performed faster without losing precision.
“This research demonstrates that AI does not just automate existing procedures; it changes the way problems are posed,” Drala said. “Instead of relying solely on faster computers, AI models can directly learn atomic motion, increasing the efficiency of the algorithms themselves. Computational chemistry and molecular physics are entering a phase where AI co-creates new scientific practices. It does not replace the laws of physics, but helps find a shorter path to their practical application.”
The study was published in the Journal of Chemical Theory and Computation (doi: 10.1021/acs.jctc.5c01689). (PAP)
kmp/ zan/
tr. RL
