DeepModeling

DeepFlame is an open-source platform for combustion fluid computation developed in the era of AI for Science, aiming to promote the application of combustion fluid simulation technology in scientific research and industry [1–4]. Since its release, the platform has attracted extensive attention from academia and industry, and continues to attract outstanding developers and user communities to participate in its construction.

In this update, DeepFlame further expands the platform's engineering steady-state computing capabilities and the flexibility of thermochemical calculations. The new version of DeepFlame provides an option for sensible enthalpy (hs) in energy transport and computation, which can be selected during solver compilation. On this basis, DeepFlame has added temperature calculation capabilities independent of Cantera, supporting local thermodynamic state solving based on Newton iteration. Users can switch with one click through environment variables during compilation to meet the integration needs of different users. In addition, the DeepFlame 1.6 version introduces the steady-state solver dfSteadyFoam, which adds component transport and chemical reactions based on rhoSimpleFoam. It is suitable for calculating various subsonic reactive flow steady-state problems, broadening the engineering application scenarios of DeepFlame. Furthermore, we have optimized the build script, which can display system architecture information after the build is completed, improving the transparency of platform deployment. At the same time, DeepFlame has been adapted and optimized based on the domestic Kunpeng platform, achieving significant performance improvements in the inference and solving parts. The documentation system has also been upgraded to sphinx-book-theme, and with the new Python and operating system environments, the documentation interface is more modern and the organization is clearer. We have also added a 2D Riemann case as a test case for various flux splitting in dfHighSpeedFoam, further enriching the verification cases for high-speed reactive flow solvers.

Version Update Overview

• Based on rhoSimpleFoam, the dfSteadyFoam steady-state compressible reactive flow solver is launched, suitable for calculating various subsonic steady-state cases.

• The energy calculation provides a sensible enthalpy (hs) option, and can complete temperature calculation as well as enthalpy and specific heat capacity solving independently of Cantera. Users can freely switch the solving mode through the environment variable CANTERA_THERMO.

• The installation script is optimized, and the system architecture is automatically displayed after the build is completed, facilitating deployment and debugging.

• The documentation system is upgraded to sphinx-book-theme, improving documentation compatibility and reading experience.

• The problem of repeated initialization of Cantera Reactor is solved by calling syncState(), achieving an acceleration effect of 8% to 10%.

• A 2D Riemann case is added as a test case for various flux splitting in dfHighSpeedFoam.

New Feature 1: Steady-State Compressible Flow Solver dfSteadyFoam and Flexible Thermodynamic Modeling

This version introduces a new steady-state solver dfSteadyFoam. Based on the SIMPLE algorithm, this solver adds multi-component transport equations and chemical reaction solving, and is suitable for steady-state calculation of compressible subsonic reactive flows. It can be widely applied to steady-state problem solving and industrial engineering scenarios, expanding the applicability of DeepFlame in engineering problems.

Figure1 Steady-state fuel field obtained from 2D non-reactive single-nozzle case calculated using dfSteadyFoam

Figure2 Steady-state temperature field obtained from 2D non-reactive single-nozzle case calculated using dfSteadyFoam

The above figures show the results obtained from the 2D non-reactive single-nozzle case calculated using dfSteadyFoam. The flow field reaches relative stability after approximately 3500 iterations, and the results are basically consistent with those calculated by the steady-state solver constructed based on reactingFoam and rhoSimpleFoam in OpenFoam, thus verifying the accuracy of the solver. At the same time, it can be found that this solver cannot capture the detailed structures of the flow field, such as the instability at the mixing of fuel and oxidizer. However, using the steady-state solver to calculate the relatively stable flow field of this cold-state case on 20 cores takes less than 3 hours, which is about 4 times faster than the transient solver. Therefore, it is more suitable for calculating large-scale engineering problems and the initial field of transient problems to accelerate the simulation.

New Feature 2: Providing Sensible Enthalpy as an Option for Energy Calculation, and Capability to Complete Temperature Calculation Independently of Cantera

In terms of thermophysical properties, the energy model of DeepFlame now supports sensible enthalpy (hs) as an option for energy transport, alongside the original absolute enthalpy (ha) and absolute energy (ea). Users can select it through the function CanteraMixture::setEnergyName in createFields.H.

In addition, users can independently complete the calculation of temperature, enthalpy, and specific heat capacity without calling Cantera through the Newton iteration method implemented in DeepFlame. Users can control the calculation method through the environment variable CANTERA_THERMO (setting it to 1 during compilation means using Cantera for calculation, and 0 means using the Newton iteration calculation implemented by DeepFlame itself), which significantly improves the portability and control flexibility of the platform.

New Feature 3: New Progress in Domestic High-Performance Adaptation

DeepFlame has completed full-stack in-depth optimization based on the domestic Kunpeng platform. Firstly, the compilation and dependency library calling links are reconstructed based on the Kunpeng high-performance suite, realizing efficient and seamless migration from x86 to ARM architecture. In addition, DeepFlame makes full use of the unique matrix acceleration unit and high-bandwidth memory architecture of the Kunpeng platform, realizes hardware-level acceleration for core operations such as dense matrix multiplication, and dynamically adjusts parameters such as BatchSize, process/thread count involved in neural network inference, greatly improving the inference performance. Finally, for the multi-threaded code in the solving part, by reconstructing the task scheduling strategy, optimizing thread pinning and resource isolation, the inter-core competition problem is effectively avoided, and the solving efficiency of each control equation is significantly improved.

Figure3 Time comparison between Kunpeng Professional Edition and NVIDIA A100 in solving and inference parts

Demonstration of Newly Added Cases

To further verify the ability of DeepFlame to resolve shock wave structures and contact discontinuities under high-speed compressible flows, this version adds a 2D Riemann problem case [4]. This problem includes multi-shock interaction and strong shear characteristics, and is an important test for the robustness and low-dissipation characteristics of high-order schemes.

In this case, a flow field with significantly different initial conditions in four regions is set in a 1 m×1 m computational domain, and a high-resolution mesh of 2000×2000 is used for solving. The initial shock intersection point triggers complex wave system evolution, including reflected shocks, Mach stems, and shear layers.

Figure4 Density contours and isopleths of Riemann problem under different numerical schemes

As shown in the figure, the comparison of simulation effects under different numerical schemes shows that the new numerical scheme has lower numerical dissipation compared with the traditional KNP scheme, and can more clearly capture small-scale structures near the contact surface, such as shear layer unstable ripples.

Other New Features and Optimizations

• To improve the user's deployment experience on multiple platforms, the installation script install.sh is optimized in this version. After the compilation process is completed, the script will automatically detect and display the system architecture used for the current build (such as x86_64, aarch64, etc.), facilitating users to quickly confirm the target platform, improving the visibility and debugging efficiency of the build process, and is especially suitable for multi-architecture cross-compilation scenarios.

• The documentation system of DeepFlame has also been upgraded. The theme is switched to the more modern sphinx-book-theme, and the documentation build environment is upgraded to Ubuntu 24.04 and Python 3.12 to adapt to the continuous development of the Python package ecosystem. The new documentation structure is clearer, supports multi-level directories and code block folding, and has stronger Jupyter compatibility, bringing users a smoother reading and learning experience.

Quick Access to DeepFlame

The GitHub repository address of DeepFlame in the DeepModeling community is:

https://github.com/deepmodeling/deepflame-dev

The release tag of DeepFlame 1.6 version:

https://github.com/deepmodeling/deepflame-dev/releases/tag/v1.6

Documentation (including installation methods, input/output parameter introduction, function introduction, case introduction, developer notes, etc.):

https://deepflame.deepmodeling.com/en/latest/

List of Feature Updates and Bug Fixes

v1.6.0

• Added steady-state compressible flow solver dfSteadyFoam by @pkuLmq in #559, which supports turbulence and efficiently calculates steady-state solutions using the SIMPLE algorithm.

• Extended the energy model by @user20250420 in #556, supporting sensible enthalpy (hs) in addition to the original absolute enthalpy (ha) and internal energy (ea). In addition, the calculation of temperature (T), enthalpy (h), and specific heat capacity at constant pressure (cp) can now be completed independently of Cantera. This change improves the flexibility of the platform and removes the dependency on external dependencies. Users can control the calculation method by setting the environment variable CANTERA_THERMO in bashrc, where CANTERA_THERMO=1 (currently default) uses Cantera, and CANTERA_THERMO=0 uses the built-in calculation method of DeepFlame.

• Optimized the installation script install.sh by @seeudong in #546, which will automatically display system architecture information (such as x86_64) after the build is completed, facilitating users to confirm the build platform. At the same time, the transplantation and optimization of DeepFlame have been carried out on domestic software and hardware platforms, achieving comprehensive performance improvement under the Arm architecture.

• Switched the documentation system to sphinx-book-theme and updated the build environment to Ubuntu 24.04 and Python 3.12 by @njzjz in #547, improving the aesthetics and build compatibility of the documentation interface.

• Solved the problem of repeated initialization of Cantera Reactor by calling syncState() by @xiao312 in #563.

• Added 2D Riemann case as a test case for dfHighSpeedFoam by @circlexiang in #560.

References

1. Mao R, Lin M, Zhang Y, et al. DeepFlame: A deep learning empowered open-source platform for reacting flow simulations. Computer Physics Communications, 291: 108842. (2023)

2. Mao R, Zhang M, Wang Y, Li H, et al. An integrated framework for accelerating reactive flow simulation using GPU and machine learning models. Proceedings of the Combustion Institute, 40(1-4), 105512. (2024)

3. Mao R, Dong X, Bai X, et al. DeepFlame 2.0: A new version for fully GPU-native machine learning accelerated reacting flow simulations under low-Mach conditions. Computer Physics Communications, 312: 109595. (2025)

4. Zhang M, Mao R, Li H, An Z, Chen ZX. Graphics processing unit/artificial neural network-accelerated large-eddy simulation of swirling premixed flames. Physics of Fluids, 36(5). (2024)

5. Chen H, Zhao M, Hua Q, Zhu Y. Implementation and verification of an OpenFOAM solver for gas-droplet two-phase detonation combustion. Physics of Fluids, 36(8). (2024)

1. Preface

ABACUS has released the 3.10 - LTS stable version and is still being continuously iterated. Many users hope to deploy the ABACUS software on their own machines to experience the computing efficiency improvement brought by ABACUS. However, compiling ABACUS in different server and workstation environments and achieving the highest computing efficiency in these specific environments always presents some challenges.

The ABACUS Toolchain is a set of bash script collections built into the ABACUS repository. It can help users compile and install the software dependencies required by ABACUS online or offline, automatically handle the environment variables of each dependency library, and quickly complete the ABACUS source code compilation process based on these dependency libraries, realizing an efficient, high - performance, easy - to - modify, and easy - to - port automated ABACUS compilation solution.

This tutorial is written based on the ABACUS Toolchain of the 2025 - 02 version. At present, the ABACUS Toolchain supports the following compilation and installation functions:

• GNU Toolchain, that is, the Toolchain method of compiling and installing ABACUS dependency libraries and the ABACUS body from scratch starting from a sufficient version of the GNU compilation suite (gcc, g++, gfortran, collectively referred to as GCC).
• Intel Toolchain, that is, the Toolchain method of compiling and installing ABACUS dependency libraries and the ABACUS body based on Intel's compiler, mathematical library, and parallel library (usually packaged in Intel - OneAPI or Intel - parallel - xe - studio).
• AMD Toolchain, that is, the method of compiling and installing ABACUS based on AMD's compiler and mathematical library, which is subdivided into GCC - AOCL Toolchain and AOCC - AOCL Toolchain.

At the same time, the ABACUS Toolchain also supports a series of advanced functions including functional plug - in support and packaged offline installation.

In general, the vision that the ABACUS Toolchain hopes to achieve is:

• To facilitate users to efficiently compile the ABACUS most suitable for the current server environment from the source code, and to quickly test the computing efficiency of ABACUS compiled by different dependency library types of Toolchain.
• To establish a standard process for ABACUS source code compilation. ABACUS developers can directly control the version and compilation method of each ABACUS dependency library in the Toolchain without having to compile and manually add various compilation options by themselves.

There has been a previous tutorial introducing how to use the GNU Toolchain to simply and directly compile ABACUS from scratch: ABACUS Installation Tutorial - Toolchain (1 - GNU). This solution has the best compatibility, but the compiled ABACUS may not be the most efficient, especially for many Intel - CPU servers configured with the corresponding Intel OneAPI suite. This tutorial will focus on how to use the Intel Toolchain to make the compiled ABACUS obtain higher performance.

Recently, the Beijing Institute for Scientific Intelligence, in collaboration with the Shanghai Institute for Creative Intelligence, the Zhu Tong research group at East China Normal University, and New York University Shanghai, etc., pre-published the latest research progress in the field of large atom models on ChemRxiv under the title "General reactive machine learning potentials for CHON elements".

This study proposes a complete workflow for systematically constructing universal chemical reaction machine learning potential energy surfaces (MLPs) in the era of large atom models. It has breakthroughly built universal reactive MLPs for elements C, H, O, and N. Through innovative data construction and hybrid training strategies, it achieves chemical reaction simulation capabilities approaching DFT accuracy. The team proposed a dynamic sampling method of "wide coverage + active learning", generating the RXN-xTB pre-training dataset composed of over 17 million non-equilibrium structures and the fine-tuning dataset RXN-xTB-AL containing 200,000 structures. Combined with pre-training and Δ-learning collaborative optimization, the hybrid training strategy enables the DPA-3-DF model to achieve an MAE of 0.51 kcal/mol in energy prediction and 0.49 kcal/mol/Å in force prediction, significantly surpassing various existing mainstream neural network architectures. Dynamic simulation verification shows that the model can accurately characterize the dynamic bond fission process of complex reactions, providing a new paradigm that balances quantum accuracy and molecular dynamics efficiency for catalytic design and reaction mechanism analysis. This research achievement marks a major leap in machine learning potential energy in the field of chemical reaction modeling, providing a feasible new path for the precise and efficient simulation of typical organic reactions and catalytic systems.

Paper link:
https://chemrxiv.org/engage/chemrxiv/article-details/684ffe583ba0887c33dad39b

This document describes how to install ABACUS agent and Google ADK in a local Linux environment (including WSL) as of June 12, 2025, and how to run a simple ABACUS calculation after startup.

Requirements: conda and ABACUS (recommended to install v3.10 LTS version) are installed locally.

Background

ABACUS is a density functional theory (DFT) software initiated by the University of Science and Technology of China, and open-sourced and co-constructed by multiple domestic teams including Peking University, Institute of Physics of Chinese Academy of Sciences, Beijing Academy of Science and Intelligence, and Hefei Comprehensive Artificial Intelligence Research Institute. Since adopting the LGPL3.0 open-source license in 2021 and further embracing the open-source sharing concept together with the DeepModeling community, it has successively released 70 iterative versions. Both the software functions and ecosystem have been significantly developed thanks to the selfless contributions of the open-source community.

Recently, the research team of Professor Chen Xiumin from the National Engineering Research Center for Vacuum Metallurgy, Kunming University of Science and Technology, in collaboration with DeepSeek, has achieved research on the microscopic reaction mechanism of vanadium removal from crude titanium tetrachloride by aluminum addition through a new method of artificial intelligence-driven scientific research (AI for Science). This study utilized the Deep Potential Molecular Dynamics (DPMD) simulation method to efficiently explore the reaction mechanism of vanadium removal by aluminum addition at the nanosecond time scale and the spatial scale of tens of thousands of atoms. Theoretical simulation analysis and experimental research show that the vanadium removal reaction is a synergistic mechanism of reduction and complexation reactions. In the Al-Cl₂-TiCl₄-VOCl₃ system, the reduction process forms polynuclear complexes with aluminum, titanium, and vanadium as central atoms bridged by Cl and O atoms. These polynuclear complexes, catalyzed by AlCl₃, convert VOCl₃ into VOCl₂ and VCl₃ through the exchange and transfer of Cl and O atoms in two reaction pathways. In this study, DPMD provides a new means to understand specific reactions from a microscopic perspective. The study of this reaction mechanism not only helps with the recycling and utilization of vanadium resources but also provides a theoretical basis and innovative ideas for the optimization and improvement of vanadium removal reagents.

We are pleased to announce to the DeepModeling community that Uni-Mol Tools [1] is officially and independently released! As an important part of the Uni-Mol ecosystem, Uni-Mol Tools provides more flexible and efficient tool support for molecular AI research and applications with its characteristics of lightweight, out-of-the-box, and scenario-oriented.

On May 6, 2025, the DeepModeling Community released Community Manifesto 2.0, planning to rapidly expand exploratory work in the field of "AI literature reading" in the near future. Today, the SciAssess project has officially joined the DeepModeling Community. Developed jointly by DeepSeek and the Beijing Academy of Scientific Intelligence, this system is a testing benchmark specifically designed to evaluate the scientific literature analysis capabilities of large language models (LLMs), aiming to advance the process of AI empowering scientific research. SciAssess will collaborate with the community to launch explorations in the field of AI for literature analysis.

Water, one of the most common yet complex molecules, has long perplexed researchers with its thermodynamic properties. A controversial hypothesis suggests that under supercooled conditions, water may exhibit a "liquid-liquid transition (LLT)"—a transformation between low-density liquid (LDL) and high-density liquid (HDL)—governed by a "second critical point (LLCP)."

However, direct observation of the LLCP has been extremely challenging. Within experimentally accessible temperature-pressure ranges, liquid water readily freezes into ice, while simulations struggle to reach the microsecond timescale. As a result, this hypothesis has long remained an unresolved "mystery".

Recently, F. Sciortino et al. leveraged the DeePMD framework with the DNN@MB-pol potential model to conduct microsecond-scale molecular dynamics simulations, achieving high precision approaching CCSD(T) calculations. For the first time, they provided strong constraints on the location of the liquid-liquid phase transition critical point in water. Published in Nature Physics under the title "Constraints on the location of the liquid-liquid critical point in water," this work opens a new chapter in understanding water's anomalous behavior.

Recently, the Journal of Chemical Theory and Computation published a research work titled DeePMD - kit v3: A Multiple - Backend Framework for Machine Learning Potentials [1]. This work focuses on a core innovation in the DeePMD - kit v3 version - the multi - backend framework. The latest version has integrated four deep learning backends: TensorFlow, PyTorch, JAX, and PaddlePaddle.