DeepModeling

Recently, the OpenLAM Team of the Beijing Academy of Artificial Intelligence for Science, Peking University, DeepModeling Technology, and the Institute of Applied Physics and Computational Mathematics have jointly launched DPA4, a new-generation model architecture tailored for the era of Large Atomic Models (LAMs). DPA4 claimed the top spot worldwide with its comprehensive performance score (CPS) on Matbench Discovery, an authoritative global benchmark for materials discovery, emerging as the latest State-of-the-Art (SOTA) model.

What makes DPA4 particularly remarkable is that it achieved this milestone not by expanding model parameters or relying on massive computing power. The previous leading model eSEN required over 300 GPU days for training, while DPA4 can theoretically reach the same accuracy with just a single consumer-grade RTX 5090 running for approximately one day. Meanwhile, its parameter count is less than one-tenth of that of eSEN.

In short, the SOTA-level accuracy that once demanded exorbitant supercomputing budgets can now be achieved using a single consumer graphics card. DPA4 is redefining the Pareto frontier between accuracy and efficiency for large atomic models.

Official Screenshot of Matbench Discovery (Data as of May 22, 2026)

Over the past two years, the DeepFlame community has witnessed the rapid development of AI for Science (AI4S) together with researchers and practitioners. Since advocating the co-construction of an AI4S open-source combustion platform in June 2022, and releasing more than twenty versions that realize full-process GPU heterogeneous solvers, we have consistently been committed to building a bridge between artificial intelligence, high-performance computing, and physical modeling.

However, in today’s era of explosive AI growth, why are many researchers’ daily routines still dominated by heavy code debugging and case configuration? True AI4S should not stop at “using AI to compute faster,” but should aim to “use AI to liberate researchers’ productivity.”

Today, we officially release DeepFlame 2.0. In this version, beyond functional updates and performance optimizations, more importantly, we formally introduce a brand-new scientific computing paradigm — AI-agent-driven scientific computing. By bringing AI agents into scientific computing workflows, researchers can leverage the power of intelligent agents to improve research efficiency and focus more on solving scientific problems themselves.

Given a chemical formula, for example Cu₁₂Sb₄S₁₃, how should the atoms be arranged in space in order to form a stable crystal? This is the problem of crystal structure prediction (Crystal Structure Prediction, CSP), one of the fundamental challenges in materials science research. Recently, the Institute of Physics, Chinese Academy of Sciences released CrystalFormer-CSP to the DeepModeling community, adopting a strategy that combines “fast thinking” and “slow thinking” to address this challenge.

In the era of AI for Science, where research workflows are being fundamentally reshaped, the ability to acquire high-throughput, high-quality data has become the key competitive edge. Yet today’s labs still face major challenges: heterogeneous hardware, closed protocols, and fragmented data flow force researchers to spend precious time on integration overhead—not scientific discovery.

An automated lab should not be a collection of expensive machines, but an extension of intelligent decision-making.

In December 2025, with the official merge of v0.10.13, Uni-Lab-OS enters version 1.0. We aim to provide a standardized digital infrastructure for the scientific community—breaking down device barriers and freeing innovation from tooling constraints.

You may not know this: a silicon wafer that looks perfectly smooth is, at the atomic scale, actually a dynamic stage—silicon atoms pair up into dimers, hydrogen atoms shuttle back and forth, and at high temperatures the surface even “pre-melts,” forming a quasi-liquid layer that resembles sweating. These microscopic behaviors directly affect the quality of chip manufacturing.

Recently, the research team led by Professor Li Pai and Professor Wei Xing from the Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, published a study in Small. For the first time, they systematically revealed how the Si(001) surface evolves under different temperatures and hydrogen environments, and captured its pre-melting phenomenon before bulk melting occurs. All of this relies on a key tool: Deep Potential (DP).

From the vivid colors of smartphone displays and the high efficiency of photovoltaic solar panels, to high–energy-density batteries and sharp bio-fluorescent imaging, organic optoelectronic molecules are indispensable. They serve as the “soul” and “modulator” of optoelectronic functions. With structural tunability at the molecular scale, they continuously enable the evolution of optoelectronic devices and their broad application scenarios.

However, to fully unlock the potential of organic optoelectronic materials, it is crucial to efficiently understand—across multiple scales—the intrinsic links between molecular structure, material properties, and device performance.

Recently, the Functional Molecular Design Team of AI for Science Institute (AISI), together with the DP Technology development team, in collaboration with Peking University, Sinopec Research Institute of Petroleum Processing, Shandong University, Henan Normal University, Shenzhen Institute of Synthetic Biology, and several other institutions, introduced OCNet—a pretraining framework for organic optoelectronic materials built upon the Uni-Mol architecture. OCNet is trained on tens of millions of conjugated molecules and their dimers.

OCNet achieves, for the first time, a unified virtual representation spanning molecules, mesoscale materials, and devices: it surpasses existing SOTA models by 20% on molecular-scale performance, enables cross-material generalizable mobility prediction in amorphous organic thin films for the first time, and delivers near-real-time, high-accuracy prediction of device-level photovoltaic efficiency. The work has been published in npj Computational Materials (doi: 10.1038/s41524-025-01788-y).

Hybrid functionals (HDFs) overcome the shortcomings of local/semi-local functionals—such as the underestimation of band gaps—by incorporating exact exchange (EXX), but this comes at the cost of high computational expense. ABACUS combined with LibRI enables linear-scaling calculations of hybrid functionals, and on this basis, applying space-group symmetry can further reduce the computational load.

Prior to version 3.8.0, ABACUS already supported symmetry acceleration for local/semi-local functionals: it reduces the number of Kohn-Sham (KS) equations to be solved by reducing k-points to the irreducible Brillouin zone (IBZ). However, due to the lack of implementation for space-group transformations of the density matrix, symmetry acceleration was not supported for cases involving non-local Hamiltonians (e.g., hybrid functionals). On the other hand, symmetry reduction can also be applied to real-space two-electron integrals (ERIs) for the EXX term. Nevertheless, currently available software (such as CRYSTAL and Turbomole) only implements this for algorithms that directly compute four-center integrals, without further accelerating symmetry application based on the resolution of the identity (RI) method—a common approach to speed up ERI calculations.

The official release of UniMol_Tools v0.15 introduces lightweight pre-training and a synchronized full-process command-line tool based on Hydra. Developers can complete the entire workflow from preprocessing → pre-training → fine-tuning → property prediction with just a few lines of code, and the reproduced results are nearly identical to those of the original Uni-Mol. This new version aims to provide an efficient and reproducible computing platform for research in materials science, medicinal chemistry, and molecular design.

In recent years, with the rapid advancement of mRNA vaccines and nucleic acid drugs, lipid nanoparticles (LNPs) have emerged as one of the most crucial drug delivery tools. However, the performance of LNPs depends on various lipid components and their proportions. Experimental optimization is not only time-consuming and labor-intensive but also struggles to cover the vast design space. Recently, the work led by Alvin Chan and his team was published in Nature Nanotechnology under the title "Designing lipid nanoparticles using a transformer-based neural network". The study proposes a transformer-based neural network model called COMET, which integrates molecular structures and formulation parameters to predict LNP performance. A key component of this process is the use of Uni-Mol as the core tool for molecular representation learning.

AI-driven drug discovery relies on the accurate characterization of molecular features. The newly launched MoleculeCLA dataset by Professor Lanyan Yan's team at the Institute for Intelligent Industry, Tsinghua University, provides a new, multi-dimensional evaluation platform for molecular representation by generating large-scale computationally docked chemical, physical, and biological properties with no experimental noise. On this benchmark, Uni-Mol performed exceptionally well in end-to-end fine-tuning, achieving an average Pearson correlation coefficient of 0.68, ranking first among all pre-trained deep models and traditional molecular descriptors. This fully demonstrates its advantages in molecular feature extraction and prediction tasks. The preprint of the research paper entitled "MoleculeCLA: Rethinking Molecular Benchmark via Computational Ligand-Target Binding Analysis" has been published on arXiv.