We are pleased to announce the release of DeepFlame v2.0, marking a paradigm shift in our journey toward Agent-driven Combustion Scientific Computing. Building on over 3 years of practices in the AI+HPC+CombustionCFD methodology, this version goes beyond the traditional goal of mere computational acceleration. By integrating AI agents directly into the workflow to automate tedious coding and simulation tasks, DeepFlame v2.0 aims to liberate productivity and empower researchers to focus exclusively on scientific innovation.

DeepFlame Agent ecosystem

1. CoCo - GPU Programming Agent

CoCo is a code-migration agent for DeepFlame-GPU that understands the semantics of legacy C++ numerical algorithms and automatically generates, reviews, and tests CUDA code following the DeepFlame-GPU framework. It significantly lowers the barrier to GPU-accelerated CFD development, allowing researchers to focus on physical modeling rather than CUDA programming.

2. FlamePilot – CFD Simulation Agent for Combustion

FlamePilot is a CFD agent designed as a “digital teammate” for combustion simulations in DeepFlame. Through natural language interaction, it assists users in setting up simulations, autonomously diagnoses issues based on runtime feedback, proposes improvements, and performs corrective optimizations until convergence is achieved.
Check https://github.com/deepflame-ai/flamepilot for more information.

3. DFODE-kit Trainer - DFODE Neural Network Training Agent

DFODE-kit Trainer enables agent-driven training of combustion chemistry DNN models. Using natural language instructions, it autonomously handles operating condition setup, data generation, model training, and validation, greatly improving the efficiency and accessibility of neural chemistry model development.
Check https://github.com/deepflame-ai/DFODE-kit/blob/agent/agent_user_guide_zh.md for more information.

The future agents will be updated in deepflame-dev/agents folder.

New features

DFODE-kit - Deep Learning Package for Combustion Kinetics

DFODE-kit is an open-source Python package designed to accelerate combustion simulations by efficiently solving flame chemical kinetics governed by high-dimensional stiff ordinary differential equations (ODEs). This package integrates deep learning methodologies to replace conventional numerical integration, enabling significant speedups and improved accuracy. To train your own DNN model, try to follow DFODE-kit tutorials or DFODE-kit Trainer agent.

Other updates since v1.6.0

• Add 2D Riemann problem case by @circlexiang in #560
• update README.md by @pkuLmq in #561 #562
• Improve Initialization in dfChemistryModel by @xiao312 in #563
• CI: Change actions to self-hosted runner by @TimoLin in #557
• Bug fix: improve the logic of sorting singular values in SIGMA sgs model. by @TimoLin in #574
• Add support for the localBlended scheme by @TimoLin in #577
• Use absolute enthalpy in df0DFoam solver by @Copilot & @TimoLin in #578

New Contributors

• @circlexiang made their first contribution in #560
• @TimoLin made their first contribution in #557
• GitHub Copilot is introduced to fix simple issues like #578

Full Changelog: v1.6.0...v2.0.0

Features

• Add a new solver, dfSteadyFoam, a steady-state compressible flow solver by @pkuLmq in #559. It supports turbulence and uses the flexible PIMPLE algorithm to efficiently compute steady solutions
• Extend energy model to support sensible enthalpy (hs) in addition to absolute enthalpy (ha) and internal energy (ea). Moreover, the calculations of T, h, and cp can now be performed independently of Cantera, utilizing the methodology introduced by @user20250420 in #556. This adjustment improves flexibility and removes external dependencies. CANTERA_THERMO is added in bashrc to control method to calculate h, cp and update T. CANTERA_THERMO=1(current default) means using Cantera and CANTERA_THERMO=0 means using deepflame.
• Improve install.sh by displaying system architecture (e.g., x86_64) after build for clearer platform information by @seeudong in #546
• Switch to sphinx-book-theme and update documentation build environment (Ubuntu 24.04, Python 3.12) for better appearance and compatibility by @njzjz in #547
• Addresses the issue of repetitive initialization in the Cantera Reactor by calling syncState() by @xiao312 in #563
• Add 2D Riemann problem example case by @circlexiang in #560

Features

• provide a new boundary condition, totalFlowRateAdvectiveDiffusion (adopted from OpenFOAM v7), which accounts for diffusion effects at the boundary by @maorz1998 in #524
• develop a new solver, dfBuoyancyFoam (adopted from fireFoam), a transient, compressible solver designed to model turbulent reacting flows while incorporating buoyancy effects by @chenzhenyi-123 in #539
• add new combustion models (including infinitelyFastChemistry and eddyDissipationModel) to the dfBuoyancyFoam solver by @chenzhenyi-123 in #538
• add radiation models by @chenzhenyi-123 in #536 and #540, which are integrated into the current dfBuoyancyFoam solver and can be incorporated into other solvers if needed
• imply DeepFlame on Arm architecture platforms by @seeudong in #541, details can be found in README

Bug Fix and Improvements

• adjust the split method of rhoY in dfHighSpeedFoam by @pkuLmq in #517

Full Changelog: v1.4.0...v1.5.0

Features

• reorganize the update order of mass, velocity and temperature for Lagrangian particles and introduce the liquidEvaporationSpalding model as new evaporation model by @pkuLmq in #504
• add source terms for liquid phase in the dfLowMachFoam solver by @pkuLmq in #506
• incorporate Euler-Lagrangian source terms into the dfHighSpeedFoam solver by @pkuLmq in #510 to facilitate numerical simulations of two-phase supersonic reactive flows
• provide new flux schemes (including HLLC and HLLCP) for dfHighSpeedFoam solver (adopted from detonationFoam ) and do some modifications @pkuLmq in #510
• add lagrangianExtraFunctionObjects function (adopted from lagrangianExtraFunctionObjects ) in submodules to write to disk in the old positions file format by @pkuLmq in #511
• introduce new cases to evaluate the accuracy of dfHighSpeedFoam solver and provide two-phase 1D/2D detonation cases by @pkuLmq in #510
• add AUSMDV scheme as new flux scheme by @risinyoung in #512
• add compatibility of neural network inference for chemical source terms with the Baidu PaddlePaddle framework by @BiteBytePKU in #500

Bug Fix and Improvements

• adjust original examples referring to the modification of solvers

Build Test

• add 2D aachenBomb case in test by @pkuLmq in #509

Document

• update PaddlePaddle options for DNN model development and inference in document homepage by @xiao312 in #501

Full Changelog: v1.3.0...v1.4.0

Features

• Added optional compile option "--use_ode_gpu_solver" to enable DeepFlame to accelerate ODE solving by GPU when ode_gpu_solver is available by @dangglxxx in #455

Bug Fix and Improvements

• Fix on coefficients of RK3 by @pkuLmq in #416

Full Changelog: v1.1.2...v1.1.3

Features

• Use diffusion correction only when turbulence type is laminar for dfHighSpeedFoam by @hanli94 in #318
• Add a 2-D temporally evolving jet flame case for dfLowMachFoam by @hanli94 in #354
• Modify torchFunction.H and update examples/df0DFoam/zeroD_cubicReactor/H2/pytorchIntegrator by @dangglxxx in #364
• Yr/see amgx by @STwangyingrui in #382
• add options for thermo DNN by @JX278 in #391
• update flareFGM by @DongWangChina in #397
• Add a 2D supercritical TGV case by @JX278 in #398
• modified to DF-ODENet v0.1 by @risinyoung in #399
• Merge GPU version by @maorz1998 and @STwangyingrui in #403
• Update DNN examples by @hanli94 in #408
• Update tutorial examples by @hanli94 in #411
• Update DeePFGM by @haixiao-stack in #400

Bug Fix and Improvements

• Update Allrun to allow root user by @LefoMount in #311
• Fix bugs when using RK2SSP and RK3SSP and update time statistics in dfHighSpeedFoam by @hanli94 in #315
• Update torchFunctions by @minzhang0929 in #383
• update pytorchFunctions by @minzhang0929 in #384
• optimize calculate_viscosity_kernel by @STwangyingrui in #387
• fix a bug: mixture fraction boundary condition by @JX278 in #407

Build Test

• add the testing network to the df0D case by @JX278 in #386
• add a test for torch by @JX278 in #406
• Update Docs and test by @JX278 in #414

Document

• Update README.md by @xiao312 in #323
• Update install document: df-notorch python=3.8 by @DongWangChina in #334
• Update install.rst by @risinyoung in #368
• Update install.rst by @hanli94 in #376
• Add support for python version agnostic building by @caic99 in #379
• Update index.rst by @hanli94 in #409
• Update README.md by @hanli94 in #410
• Update README.md by @xiao312 in #413
• Update the citation by @JX278 in #415

New Contributors

• @LefoMount made their first contribution in #311
• @DongWangChina made their first contribution in #334
• @dangglxxx made their first contribution in #364
• @risinyoung made their first contribution in #368
• @caic99 made their first contribution in #379

Full Changelog: v1.2.0...v1.3.0

Features

• add the validated cuda kernels for ueqn by @STwangyingrui in #239
• introduce new class dfMatrix and construct ueqn matrix with csr format by @maorz1998 in #240
• Gpu by @maorz1998 in #244
• GPU version of YEqn by @maorz1998 in #281
• GPU version of EEqn by @STwangyingrui in #283
• GPU version of UEqn.A and UEqn.H by @maorz1998 in #284
• primary implementation of correctBoundaryConditions on GPU by @maorz1998 in #291
• add fixedValue boundary, debugging by @maorz1998 in #301
• add two submodules by @JX278 in #302
• Update compressibility calculation for real fluid in CanteraMixture.H by @wengzf20 in #303
• Gpu by @maorz1998 in #307
• realize RK2SSP and RK3SSP in dfHighSpeedFoam and add wrate option for chemistry calculation by @pkuLmq in #309
• fixed value by @maorz1998 in #305
• recent modification by @maorz1998 in #253
• initial commit of turbulence->divDevRhoReff(U) by @maorz1998 in #255

Bug Fix and Improvements

• rhoEqn & fix bugs in UEqn by @maorz1998 in #277
• fix bugs in 1.2 by @maorz1998 in #308
• update PaSR by @minzhang0929 in #300
• add calculation of net production rates in dfChemistryModel by @pkuLmq in #304
• DeePFGM-v1.0 by @haixiao-stack in #312

Refactoring

• preliminary refactoring by @maorz1998 in #272
• remove rebundant data transfer between equetions and construct bounda… by @maorz1998 in #286
• several refactors of GPU branch by @STwangyingrui in #289
• optimez gpu yeqn: merge species by @STwangyingrui in #292
• eliminate redundant operations when using UnityLewis to enhance perfo… by @maorz1998 in #294
• update docs, configues, and installation by @maorz1998 in #306

New Contributors

• @STwangyingrui made their first contribution in #239

Full Changelog: v1.1.0...v1.2.0

Features

• fix energy equation and change convection term in dfHighSpeedFoam, add new example for dfHighSpeedFoam by @pkuLmq in #287
• add DeePFGM features by @haixiao-stack in #297

Bug Fix and Improvements

• remove time index output by @wengzf20 in #282

Build Test

• add the test for dfHighSpeedFoam by @JX278 in #288

Refactoring

• Update EEqn.H, change alphaEff to alpha for laminar by @zhixchen in #293
• Update YEqn.H, rearrange laminar bool by @zhixchen in #295

Full Changelog: v.1.1.4...v1.1.5

Features

• support absolute enthalpy/energy in CanteraMixture by @OpenFOAMFans in #267
• Create access to the reaction rate of each species involved in a given reaction by @minzhang0929 in #280

Bug Fix and Improvements

• catch error from Cantera and report in OpenFOAM by @wengzf20 in #268
• Convert XML formatted mechanism files to YAML format by @xiao312 in #269
• change "hexRef8" to "hexRef8of" to avoid conflict by @pkuLmq in #279

Build Test

• compile to different folders for opt/dbg by @OpenFOAMFans in #270

Document

• Update /mechanisms/H2/ReadMe.md by @xiao312 in #271

Full Changelog: v1.1.3...v.1.1.4

Features

• Change psiThermo to rhoThermo by @wengzf20 in #256
• Update rho_ in rhoThermo from flareFGM by @wengzf20 in #260
• check whether mechanism exist before Cantera read it, to make error more readable when mechanism file not exist by @OpenFOAMFans in #261

Full Changelog: v1.1.2...v1.1.3