Learning molecular dynamics: MDtrajNet

Back in March 2022, we introduced a novel concept of directly learning dynamics via 4D-spacetime atomistic AI models (4D models for short). The idea is to predict the nuclear coordinates as a continuous function of time. The model GICnet was published in JPCL in 2023:

  • Fuchun Ge, Lina Zhang, Yi-Fan Hou, Yuxinxin Chen, Arif Ullah, Pavlo O. Dral*. Four-dimensional-spacetime atomistic artificial intelligence models. J. Phys. Chem. Lett. 2023, 14, 7732–7743. DOI: 10.1021/acs.jpclett.3c01592.

However, this proof-of-concept work is not perfect and has many limitations. One major problem is that the GICnet model is not generalizable, i.e., it can only be trained and used for a specific molecule. Now we get a better choice, MDtrajNet, which overcomes these limitations. We also present MDtrajNet-1, a foundational model that directly generates MD trajectories across the chemical space.

MDtrajNet

MDtrajNet combines equivariant neural networks with a Transformer-based architecture to achieve strong accuracy and transferability in predicting long-time trajectories for both known and unseen systems. The errors of the trajectories generated by the foundational model MDtrajNet-1 for various molecular systems are close to those of the conventional ab initio MD. The model’s flexible design supports diverse application scenarios, including different statistical ensemble, boundary conditons, and interaction types.

See our preprint for more details:

  • Fuchun Ge and Pavlo O. Dral*. Artificial intelligence for direct prediction of molecular dynamics across chemical space. ChemRxiv. 2025. DOI: 10.26434/chemrxiv-2025-kc7sn.

备注

In this tutorial, we only talk about the Python API. Currently, the usage of MDtrajNet via input file/command line is not supported.

Now, let’s see how to use MDtrajNet in MLatom!

Prerequisites

  • MLatom 3.17.4 or later

  • e3nn 0.4.4 (no guarantee for other versions)

备注

MLatom will download MDtrajNet-1 models for you. If the download fails, you can download it by yourself by following the error message.

Tutorial

Get started with examples on how to use it (notebook file and model file)

mdtrajnet