DP Technology's scientific loop is strongest in the middle

DP Technology founder and CEO Sun Weijie. The company he co-founded in 2018 is trying to expand a molecular-simulation core into a full scientific workflow. Photograph supplied by the interviewee to National Business Daily.[6]

Most AI companies begin with a model and search for places to put it. DP Technology began with a harder constraint: if software proposes an answer about a molecule or material, physics and experiment eventually get a vote.

That origin makes the Beijing-based company a useful AI-China dossier. By July 2026, DP presented itself not merely as a model lab or scientific-software vendor, but as a supplier for the entire research cycle: finding literature, running computation, designing candidates, connecting laboratory work, and feeding results back into the next decision.[1] The ambition is coherent enough to take seriously. The evidence, however, is not equally strong at every layer.

The computational middle of the stack is public and inspectable: open-source code, versioned releases, documentation, and peer-reviewed work. The newer reading and agent products are visible as software but described mostly through company metrics. The experimental edge is tangible—a product page specifies synthesis, assays, and laboratory feedback—yet the public record does not show that every part already operates as one autonomous scientific system.[2][3][4][5] DP's significance lies in that gap. It is trying to turn a credible technical core into a much larger product before the phrase “AI scientist” has acquired a settled meaning.

The Company Started With The Simulator

DP says it was founded in 2018, with offices and research-and-development centers across Beijing, Shanghai, Shenzhen, and Yibin. Its founders are CEO Sun Weijie and chief scientist Zhang Linfeng; mathematician E Weinan serves as chief scientific adviser.[1] An independent 2023 profile traced the business to “Deep Potential,” the approach behind its name: machine learning is used to approximate the potential-energy surface that governs atomic interactions, making some molecular simulations far less expensive than repeatedly calculating them from first principles.[7]

This is not a language model making a plausible guess about chemistry. A learned interatomic potential predicts quantities such as energy and force so that a molecular-dynamics system can advance atoms through time. The useful output is therefore constrained by conservation, symmetry, training data, numerical stability, and the physical regime in which the model is applied. A model can still be wrong, but the failure is measurable in a domain where a fluent paragraph is irrelevant.

The best evidence for DP's model strategy is DPA-2, published in npj Computational Materials on December 19, 2024. Its multi-task pre-training setup combined 18 datasets covering 73 chemical elements, while retaining separate fitting networks for datasets produced under different quantum-mechanical settings. On the paper's stated test setup, the authors reported lower weighted-average zero-shot energy and force errors than both a pre-trained MACE comparison and a single-task DPA-2 baseline.[3]

Those are bounded results, not a universal certificate for matter. The paper itself says its pre-training data were deficient for areas including two-dimensional materials and calls large atomic models a long-term undertaking.[3] That qualification is important. DP's strongest scientific artifact does not claim that one model has solved chemistry; it shows a way to reuse representations across mismatched datasets, then fine-tune and validate them for a narrower task.

Open Software Makes The Middle Legible

The second credibility signal is maintenance. On March 19, 2026, the DeepModeling project released DeePMD-kit v3.1.3. The stable release added a built-in path for downloading pre-trained models, introduced DPA3-Omol-Large through that mechanism, expanded distributed PyTorch training, and continued work on an experimental exportable PyTorch backend.[2] These details sound mundane beside an “AI scientist” pitch. They are actually more revealing.

Scientific models become useful through packaging: supported backends, reproducible configurations, diagnostics, distributed training, model export, and a way to move a trained potential into an existing simulation workflow.[2] That is the difference between publishing a model idea and maintaining an instrument other researchers can place in a larger apparatus.

Open code does not prove that DP's commercial products are open, nor does a release note establish independent scientific performance. It does establish an inspectable engineering lineage beneath the company. Users can see what changed, which backends are supported, where experimental components remain experimental, and how a model is invoked. In a sector crowded with closed demos, that maintenance trail is strategic capital.

Reading, Computing, And Experiment Are Becoming One Sales Surface

DP's current company page groups its products under a “Particle Universe” engine and a Science-as-a-Service model. The portfolio spans Bohrium Science Navigator for literature and research discovery, Bohrium Lebesgue for scientific computing, domain products such as Hermite and Piloteye, Bohrium Uni-Lab for laboratory integration, and a SciMaster scientific agent.[1] The list is broad, but it has a discernible grammar: read, compute, decide, test, repeat.

The reading layer grew quickly in product terms. DP's own February 10, 2026 review says Bohrium Science Navigator launched in March 2025 and that the workspace then expanded across search, writing, knowledge management, and research tools. It claims an index of 170 million papers, 140,000 journals, 20 million scholars, and more than 50,000 research tools and AI models.[4] Those figures describe catalog scale, not accuracy, active use, or research outcomes. Still, they show why DP is building a reading interface: a scientific agent needs access not only to prose, but also to traceable papers, specialist software, datasets, and the computational environment where an answer can be checked.

At the opposite end, DP advertises a wet-lab capability for hit discovery. Its product page describes cell-biology assays for screening and confirming compounds, synthesis support, and a loop in which experimental feedback refines models and design strategy.[5] This is the most consequential layer and the least reducible to an ordinary AI demo. A molecule cannot be willed into existence by a benchmark score. It has to be synthesizable, testable, active in an assay, and worth advancing after failure modes appear.

The important company claim is therefore not that one agent can do everything. It is that DP can own enough adjacent handoffs to reduce the loss between them. Literature evidence becomes a computational question; a model becomes a candidate; a candidate becomes an experiment; an experimental result becomes new data. If provenance survives those transitions, the loop can improve. If each product remains a separate tab connected by consultants and manual file transfer, the portfolio is only a bundle.

The Vertical Stack Is Both Moat And Risk

Sun's public language has been expansive for years. In a Chinese-language interview published by National Business Daily on January 1, 2025, he argued that scientific work would move toward cloud delivery and compared AI-for-science tools with machinery that makes exploration less dependent on blind trial and error.[6] TechCrunch's earlier profile showed the commercial logic underneath that vision: DP sold scientific-computing software, domain products, and tailored research services to customers that could not use the tooling alone.[7]

Services are not a footnote here. They reveal the difficulty of the market. A researcher-facing literature interface, a high-performance molecular simulator, a drug-design application, a laboratory orchestration layer, and an assay operation serve different users, carry different validation burdens, and fail on different clocks. The more of that chain DP owns, the more data and workflow context it can retain. The more of it DP owns, the more integration work it must also absorb.

That makes DP different from a general-model vendor that can expose an API and let customers assemble the rest. Scientific work has expensive ground truth. A failed generation can be discarded in seconds; a failed compound may consume weeks of synthesis and testing. A material candidate may perform well in simulation and still collapse under manufacturing conditions. The company can create a moat only if its loop learns from those expensive failures without hiding uncertainty or overfitting to one customer's laboratory.

The stack also creates a governance question. Scientific agents should not merely return an answer; they should preserve the paper, dataset, model version, simulation settings, code environment, candidate history, and assay result that produced it. DP's product map points toward such continuity, but the public pages cited here do not yet provide a complete, independently audited provenance standard spanning the full portfolio.[1][4][5] “Closed loop” is therefore a direction of travel, not a verified state.

What Would Prove The Dossier

Four signals would make DP's full-stack thesis more convincing.

First, the company could publish prospective outcomes: candidates selected before an experiment, the number that survived each stage, the baselines they beat, and the time and cost consumed. Retrospective case studies are useful, but a closed loop is best tested when failure cannot be edited out afterward.

Second, product-level provenance should become visible. A Bohrium result ought to identify the literature snapshot, tool version, model artifact, parameters, compute environment, and laboratory handoff behind it. Replaying that chain matters more than giving a scientific agent a personable interface.

Third, external teams should reproduce claims across laboratories and problem classes. DPA-2's paper provides a clear evaluation boundary and even names a coverage weakness.[3] The broader product suite needs the same habit: what was tested, against which baseline, in whose laboratory, with which hardware, and where the system should not be trusted.

Fourth, the portfolio must show that integration creates value rather than lock-in. Exportable data, callable tools, stable interfaces, and support for outside models would indicate that DP is selling a scientific workflow rather than a sealed scientific cloud. The open DeepModeling layer makes that possibility credible; it does not guarantee the commercial stack will follow the same pattern.[2]

The falsifier is straightforward. If customers adopt DeePMD or one specialist product while the “read-compute-experiment” cycle remains mostly bespoke service work, then DP is a strong scientific-software company with an oversized platform story. If the handoffs become repeatable, traceable, and independently validated, the company will have built something rarer: an AI business whose central product is not a model response, but a scientific loop that can learn from reality.

cronfeed.work