An atomic model becomes useful when a scientist can afford to let the atoms move. In an October 30, 2026 interview published by Beijing's science commission, DPA researcher Duo Zhang described an ambition to bring accurate simulation within reach of smaller research groups. The collaboration behind DPA4 links the AI for Science Institute in Beijing, Peking University, DP Technology and the Institute of Applied Physics and Computational Mathematics.[1]
That ambition now has several kinds of evidence: an accuracy-focused model, a design built for fast simulation, and an independent extension that adds more physics. Read together, they show an AI research program changing the practical size of an experiment. They also reveal why an impressive atom count needs two companions: how long those atoms can be followed, and which interactions the model includes.
Accuracy and speed take different routes
The DPA4 preprint, first submitted in June and revised on August 20, 2026, tackles the expense of learning atomic interactions. It learns an energy function and derives forces from it, keeping the two quantities mathematically connected. Its architecture respects how physical quantities should behave when a configuration is rotated. The authors report results on inorganic-material and molecular datasets, with a compiled training implementation that reduces the cost of that energy-to-force calculation.[2]
These are learned approximations to reference calculations. A useful model can save the repeated expense of performing those calculations during a simulation; the quality of its predictions still depends on the configurations it encounters. The DPA4 paper reports separate benchmark settings, rather than demonstrating that one saved model solves every chemical problem.[2]
DPA4C pursues a complementary goal: make each simulation step cheap enough to repeat at scale. Han Wang's September 17 Peking University seminar announcement places the accuracy and deployment work alongside the broader effort to connect atomic models with materials experiments.[3] That connection is the scientific destination. The faster calculation is a means of reaching it.
Design the work that can disappear
The revealing feature of DPA4C is what happens after training. DeepModeling's August 23 account explains that learned functions associated with an atomic pair depend on its separation and element types. This restricted form lets deployment replace repeated function evaluation with an interpolation table and cached coefficients. The model is designed from the outset to admit that substitution.[4]
Memory receives the same attention. Learned per-atom states are processed in bounded chunks, and the local interaction calculation avoids sending layers of learned features between GPUs. The machines exchange atomic geometry and physical outputs instead. This changes the storage and communication burden that grows with a simulation.[4]
The saving comes from choosing a representation that can be executed economically. A table represents learned behavior; it does not supply missing chemistry.
The software documentation makes the idea tangible. In DeePMD-kit v3.2.0, training, freezing and compression are separate steps. The frozen intermediate and compressed deployment archive serve different purposes. The documented compressed route requires CUDA, the exportable PyTorch backend and single-precision model settings.[5]
A larger box has its own clock
On periodic diamond-carbon systems at 300 K, the DPA4C preprint reports 2.048 billion atoms on 1,024 16-GB NVIDIA V100 GPUs. Its 83.3–91.2% efficiency is weak scaling: atom count grows with GPU count.[6]
Keeping roughly two million atoms fixed across that allocation yields only 6.6–15.7% efficiency. Adding processors eventually leaves too little local work to offset communication.[6]
Time remains scarce. At the billion-atom endpoint and a one-femtosecond timestep, Nano advances 0.324 simulated nanoseconds per day; Plus manages 0.037. Some empirical potentials remain faster; they are computational references, not accuracy-matched baselines. On OMol-0 held-out compositions, DPA4C also trails DPA4 more substantially than on materials benchmarks.[6]
A larger box may reveal spatial patterns. Capturing a slow event also requires a sufficiently long trajectory. Capacity and elapsed physical time answer different scientific questions.
The next extension adds distance
A further signal arrived on October 16. Researchers at the National University of Singapore posted a preprint describing DPA4C-LR, which adds long-range electrostatics and electrical response to the local model. Their motivation is physical: interactions between charges can extend beyond the neighborhood a strictly local model sees.[7]
The team separates a global electrostatic calculation from local derivatives of learned charges. Its demonstrations include water, a relaxor ferroelectric, an amorphous electrolyte and a thermal interface. These are the extension's own tests, not results established by the original DPA4C scaling run. The work is an international reuse of a Chinese-origin architecture, and remains a preprint.[7]
Deployment is equally specific. The documented fast DPA4C route combines a compressed archive with GPU-enabled LAMMPS/Kokkos and inference settings captured during export. Loading an archive through a path that transfers work back to the host each step can lose the intended advantage.[5]
The emerging value is therefore a research tool that other teams can examine, run and extend. My reading of these signals is that the most consequential progress lies in making a scientific question computationally affordable while retaining a clear account of the physics. For DPA4C, the next persuasive result will connect a tractable simulation to a material behavior that researchers can independently test.