Roundup of 2025

The end of the year is a good point to pause and look back on what happened in the past year.

For me, it was certainly one of the busiest years, full of events. We have got two new babies in my research group, and one of them is mine, making 2025 a very special year for me. And of course, we also have several new important brainchildren, which shaped this year’s research focus—on that later.

Looking at the numbers, 2025 was really productive in terms of research output measured in papers: we published 11 papers and posted 8 new preprints. See the full list of publications and screenshots below. Behind these numbers is the hard work and creativity of my team members and collaborators. I am also pleased that my citation count per year keeps steadily growing and now reaches almost 2k in 2025. This is quite an unusual citation history compared to most of my peers and reflects several factors: first, the fact that many of my research ideas are only appreciated much later and then see explosive citation growth; second, after I joined Xiamen University, I was lucky to enjoy strong support that enabled me to build a productive research program and steadily increase our output.

The research we published in 2025 reflects the diversity of our interests. We have continued to work on a new direction in autonomous atomistic simulations with AI agents. If you missed it, our Aitomia is deployed online based on the Aitomistic Lab, which is another brainchild of our team. We also achieved important breakthroughs in excited-state modeling with the first universal excited-state potential, OMNI-P2x, which enables nonadiabatic dynamics for QM methods without analytical gradients (after fine-tuning). OMNI-P2x itself is based on the all-in-one learning strategy that finally got published in JCTC after spending one year as a preprint. In addition, we introduced a new approach for accurate learning of nonadiabatic couplings and extended MLatom to various fewest-switches surface hopping algorithms.

We also published a continuation of our AIQM-series methods (AIQM2 as an issue cover in Chemical Science, efficient IR spectra simulations with AIQM2, and a preprint on AIQM3), which are robust, universal potentials based on a delta-learning approach. Delta-learning continues to be a major tool in our group, and, for example, we recently showed that it can slash the required number of active-learning iterations and training points.

We finally managed to introduce our first attempt at a foundational AI model for direct prediction of MD trajectories, circumventing step-by-step propagation. It is still under peer review, but the code is already openly available in MLatom.

Speaking about MLatom, our main ML ecosystem for computational chemistry, it saw 13 releases in 2025. These included many of the above models, but also popular third-party universal ML potentials such as MACE-OFF and AIMNet-2, as well as native implementations of quality-of-life features such as IRC.

Luckily, I was awarded the generous Research Fund for International Senior Scientists by NSFC to boost our research over the next couple of years.

Besides publications, in 2025 I enjoyed meeting many amazing scientists and aspiring young researchers at eleven conferences and workshops, where I was invited to give lectures. My travel brought me to Poland, Malaysia, and many beautiful cities in China. I also finally visited the iconic Terracotta Warriors.

In 2025, we have also put lots of our effort into bringing the cutting-edge research platforms for atomistic simulations powered by AI together with Aitomistic. Aitomistic Lab is, as one of the users said, “one of the most beautiful platforms [for atomistic simulations online] I have ever seen.” We used it daily in our research and teaching, and it has sped up our research. The platform makes remote collaborations very enjoyable, too, and is one of the reasons for the high productivity of our research. Interestingly, new students start doing their projects directly by interacting with our AI agents on the platform.

Next year, 2026, is poised to be equally exciting with all the lined-up series of projects advancing our online platforms, agentic systems, universal machine learning models, and their applications, and so much more.

Happy New Year!