⏱️ Lectura: 10 min
DP Technology, the Chinese startup behind the open source software DeePMD-kit, closed a Series C round of $114 million to expand its artificial intelligence platform applied to molecular simulation. The company, part of the DeepModeling initiative, develops deep learning methods that speed up quantum chemistry calculations without sacrificing accuracy.
📑 En este artículo
The funding arrives amid a wave of investment in AI applied to science: Washington announced a $5 billion plan for AI-assisted science and startups like Axiomatic AI raised $18 million for verified engineering infrastructure. DP Technology joins that trend with a specific focus: replacing costly physics and chemistry simulations with models trained on quantum data.
TL;DR
- DP Technology closed a $114 million Series C round for its AI platform applied to science.
- The company develops DeePMD-kit, open software for molecular dynamics with Deep Potential models.
- Deep Potential models are trained on DFT data and aim for quantum accuracy at the cost of a classical simulation.
- The round arrives as the U.S. allocates $5 billion to AI for science and Axiomatic AI raises $18 million.
- DP Technology also runs Bohrium, its scientific computing cloud for training and running these models.
- The software integrates with LAMMPS and other molecular dynamics engines used in materials science and pharmaceuticals.
What Happened With DP Technology
The Series C round, reported by Yahoo Finance, puts $114 million in fresh capital into DP Technology to accelerate development of its artificial intelligence platform for materials science, computational chemistry, and drug discovery. The company is part of DeepModeling, a research collective that brings together developers of open source scientific software.
DP Technology is best known for DeePMD-kit, a package that implements the Deep Potential Molecular Dynamics (DPMD) method: a technique that trains neural networks to reproduce the energies and forces that a density functional theory (DFT) calculation would yield, but at a fraction of the computational cost. The company’s stated goal is for researchers to be able to simulate systems of millions of atoms with near-quantum accuracy.
The company also runs Bohrium, a cloud platform where users train models, run simulations, and share scientific workflows without maintaining their own clusters. Part of this round’s capital is aimed at scaling that infrastructure.
Context and History
The Deep Potential method emerged from a classic problem in computational physics: DFT calculations are accurate but extremely slow, while classical molecular dynamics with fixed force fields is fast but loses accuracy in chemical reactions, phase changes, or new materials. DeePMD-kit emerged as one of the first practical implementations to close that gap using neural networks.
The original team behind the method is linked to mathematician Weinan E, known for his work at the intersection of partial differential equations and deep learning. DP Technology commercialized and extended that academic research into a platform with support, managed cloud training, and pretrained models for different families of materials.
Since its launch, the DeePMD-kit project has remained free software on GitHub, under the DeepModeling umbrella, which also includes tools like ABACUS for DFT and dpdata for simulation data management. That open foundation explains why universities and materials labs adopted it before a commercial product existed around it.
The round also reflects a geographic divergence in how AI for science gets funded: while the United States channels public money toward initiatives like the $5 billion one, DP Technology built its user base first with free software and only later sought private capital to scale the business around Bohrium.
Technical Details and Performance
A Deep Potential model doesn’t replace physics: it learns to imitate it. The typical workflow gathers a set of atomic structures with their energies and forces calculated by DFT, trains a neural network to predict those values from each atom’s local configuration, and then uses that trained model within a standard molecular dynamics engine.
flowchart TD
A["Reference DFT calculations"] --> B["Dataset of energies and forces"]
B --> C["Deep Potential network training"]
C --> D["Trained model (.pb graph)"]
D --> E[("Molecular dynamics in LAMMPS or i-PI")]
subgraph Bohrium
C
D
endThe architecture used by DeePMD-kit describes each atom’s local environment, that is, its neighbors within a cutoff radius, and transforms it into a feature vector before passing it through a fully connected network. That design lets a single model scale from dozens to millions of atoms, because the evaluation of each atom depends only on its neighborhood, not the full system.
| Method | Accuracy | Computational cost | Typical scale |
|---|---|---|---|
| DFT (ab initio) | Reference, high | Very high, grows fast with atom count | Hundreds of atoms, picoseconds |
| Classical molecular dynamics | Depends on force field, limited in reactions | Low | Millions of atoms, microseconds or more |
| Deep Potential (DeePMD-kit) | Close to DFT within the trained domain | Medium, scales nearly linearly | Millions of atoms while keeping quantum accuracy |
💭 Key point: the same model that reproduces the physics of a DFT calculation then runs inside a classical molecular dynamics engine like LAMMPS, without invoking DFT again at each step.
The standard validation of a Deep Potential model compares, on a test set separate from training, the error between predicted energies and forces and those calculated by DFT. A model is considered production-ready when that error falls below the tolerance the researcher defined for their specific system, not a universal threshold.
⚠️ Careful: a Deep Potential model trained for a specific material or molecule doesn’t automatically generalize to different chemical systems; it requires its own DFT dataset and its own training.
That specificity is the main limitation of the approach: a model trained for a metal oxide won’t work to simulate a protein, and generating the initial DFT dataset still requires the same expensive computation the method aims to avoid at the production stage. That’s why DP Technology combines the open software with datasets and pretrained models on Bohrium, so not every user has to run that stage from scratch.
How to Start Using DeePMD-kit
DeePMD-kit installs with pip or conda and runs on both CPU and GPU. The fastest way to try it is with the package that includes GPU support:
pip install deepmd-kit[gpu]
dp --versionThat command installs the dp CLI, which centralizes model training, freezing, and evaluation. Alternatively, DeepModeling’s conda channel also installs LAMMPS with the Deep Potential plugin already compiled:
conda create -n deepmd deepmd-kit lammps -c deepmodeling -c conda-forge
conda activate deepmdTraining a model requires a JSON configuration file that defines the network descriptor, the fitting net, and where the training data lives:
{
"model": {
"type_map": ["O", "H"],
"descriptor": {
"type": "se_e2_a",
"rcut": 6.0,
"rcut_smth": 0.5,
"sel": [46, 92]
},
"fitting_net": {
"neuron": [240, 240, 240],
"resnet_dt": true
}
},
"training": {
"training_data": {
"systems": ["./data/water/training"]
},
"numb_steps": 1000000,
"seed": 1
}
}With that file saved as input.json, the full workflow is to train, freeze the model into a graph, and run an evaluation against the validation set:
dp train input.json
dp freeze -o graph.pb
dp test -m graph.pb -s ./data/water/validationThe dp test command prints the energy and force error against the validation set: it’s the concrete way to confirm the trained model is ready before using it in a production simulation, instead of just assuming it.
Impact and Analysis
DP Technology’s funding confirms that the AI-for-science category attracts capital outside of large language models. Unlike a chatbot, a Deep Potential model doesn’t generate text: it predicts a physical quantity (energy or force) that must be consistent with the laws of quantum mechanics, and its usefulness is measured against a reference DFT calculation, not against a user’s taste.
That approach has concrete applications in battery design, catalysis, semiconductor materials, and drug discovery, areas where simulating large systems with quantum accuracy over long timescales was previously prohibitive in computing cost. The existence of a dedicated cloud like Bohrium also lowers the barrier to entry: a lab without its own cluster can train and run these models by renting compute on demand.
The contrast with the rest of the AI market is notable. While MIT Technology Review reports that AI agents for science still need more reasoning than data to become reliable, DP Technology is betting on a different path: specialized models, trained against an exact physical reference, instead of generalist agents that consult literature and propose hypotheses.
What’s Next
With fresh capital, DP Technology will likely expand the catalog of pretrained models on Bohrium and the range of chemical domains covered by DeePMD-kit, currently focused on inorganic materials, water, catalysts, and some organic molecules. The broader trend of governments and private funds channeling money into AI applied to hard science will probably keep adding similar rounds in coming months, as already happened with the $5 billion initiative announced in the United States.
For the open source community, the point to watch is whether DP Technology’s commercial growth keeps DeePMD-kit’s development as free software under DeepModeling, or whether it starts reserving new features for Bohrium’s paid offering.
📖 Summary on Telegram: View summary
Try it yourself: install DeePMD-kit with pip install deepmd-kit[gpu] and run the liquid water example from the official GitHub repository to see a Deep Potential model training in minutes.
Frequently Asked Questions
What is DP Technology?
It’s a Chinese materials science and artificial intelligence startup, part of the DeepModeling initiative, known for the DeePMD-kit software and the Bohrium scientific cloud.
What is a Deep Potential model?
It’s a neural network model trained to predict the energy and forces of an atomic system with accuracy close to a DFT calculation, but at much lower computational cost.
Is DeePMD-kit free?
Yes, it’s open source software available on GitHub under DeepModeling; Bohrium, the cloud for training and running it without your own infrastructure, is a separate commercial service.
How does it differ from a language model like GPT or Claude?
It doesn’t generate text or reason in natural language: it predicts an exact physical quantity (energy, force), and its quality is measured against a reference quantum calculation, not against human preferences.
What is it used for in practice?
To simulate materials, catalysts, batteries, and some molecules at scales of millions of atoms that would be unfeasible to calculate with pure DFT.
How much capital did DP Technology raise in this round?
$114 million in a Series C round, according to Yahoo Finance.
References
- Yahoo Finance: coverage of DP Technology’s $114 million Series C round.
- GitHub, deepmodeling/deepmd-kit: official repository of the Deep Potential Molecular Dynamics software.
- MIT Technology Review: analysis of the current limits of AI agents applied to science.
- NPR: coverage of the United States’ $5 billion plan for AI applied to science.
📱 Enjoy this content? Follow @programacion on Telegram for daily tech content in Spanish: quick summaries, fresh content every day. @programacion
Imagen destacada: Foto de Google DeepMind en Unsplash
0 Comments