All activities of the official OpenFOAM special interest group (SIG) on data-driven modeling are currently organized in this repository.

The following documents contain curated lists of contents related to data-driven modeling and OpenFOAM. You are very welcome to extend these lists.

• research projects
• software projects
• educational material

Meetings take place about every two month. Get in touch if you would like to participate. A list of previous and upcoming meetings and topics is available on the SIG's wiki page.

For us, data-driven modeling comprises a variety of different topics, some of which are listed below:

• machine learning (ML): combining CFD and ML, e.g., using ML models in CFD or deriving ML models from CFD data
• data science: analyzing CFD data to guide modeling and decision making
• data engineering: aggregating, storing, and processing CFD data

The following list briefly summarizes our current objectives. The technical roadmap below details how we plan to achieve our objectives.

• short term objectives
  • reduce the technical barrier to get started with data-riven modeling in OpenFOAM to enable more people to include data-driven workflows in their applications or research
  • promote data-driven techniques that are already available in OpenFOAM, e.g., dynamic mode decomposition (DMD)
  • promote third-party library for data-driven modeling that are based on or built for OpenFOAM
• long term objectives
  • aid the understanding of when and how to use data-driven modeling in the CFD workflow
  • accelerate developments and applications of data-driven approaches around OpenFOAM
  • establish tested data-driven techniques as a natural element of CFD simulations to improve accuracy and/or speed

We would like to have short virtual meetings (max. 1h) every two month and meet in person, if possible, at the OpenFOAM conference and workshop. The bi-monthly meetings serve to update all members about the recent progress and to coordinate upcoming efforts. Everybody is welcome to join our meetings.

This section lists several steps to achieve the objectives outlined before. The list is mostly motived by the initiator's previous experience and current projects, and presents by no means a comprehensive list of all possibilities for data-driven modeling with OpenFOAM.

• Integration of the PyTorch C++ API libtroch with OpenFOAM; libtorch can be used to execute existing PyTorch ML models in OpenFOAM but also to implement the entire ML workflow in C++ (data-processing, training, hyperparameter tuning, ...)
  • adjustment of the default OpenFOAM C++ standard, e.g, C++11 -> C++14
  • third-party install script to compile and set up libtorch
  • simplify construction of standard neural network architectures with runtime-selectable components
  • adapters for OpenFOAM field data to work with
    • full geometric fields
    • boundary fields (!)
    • sub-sets of fields
  • normalization of data as part of ML pipeline
  • base class adapters with support for ML models, e.g.,
    • boundary conditions
    • Lagrangian models
    • pressure-velocity coupling
    • ...
  • high-performance enabled computations with OpenFOAM and libtorch
    • distributed training with geometric field data
    • offload workloads to GPU(s)
• Develop workflows for Hyperparameter Tuning
  • Bayesian Optimization (!)
  • RTS for activation functions (!): addressed by PiNN team
• Provide an accessible interface between OpenFOAM and Python; Python is the most widely used programming language for data analysis and ML; therefore, there should be an easy way to access OpenFOAM data in Python; there exist a number of Python packages to access OpenFOAM data, but for all of them at least one of the following points applies: only partial access, bad or missing documentation, overly complicated access to data, no active development or support, not applicable to large data (too slow/inefficient); a few options that could be explored are
  • improve documentation of existing packages
  • continue development of packages with partial access
  • create Python bindings for selected OpenFOAM classes
• compile and create documentation and tutorials
  • OpenFOAM tutorials demonstrating how to enable ML with OpenFOAM
  • open-source research projects and proposals around OpenFOAM and ML
  • blog posts, presentations, videos
  • lecture material
• organize an OpenFOAM + ML hackathon once a year

Provide PiNN solvers in OpenFOAM for different PDE systems, use those to generalize the approach into an OpenFOAM/torch API.

1. Generalize the construction of PiNN differential operators into an OpenFOAM/torch API to lower the barrier of constructing PDE (systems).
2. Generalize the construction of initial/boundary condition losses in the PiNN PDE (system) solution.

• Extend the potential flow PiNN with OpenFOAM's boundary conditions: zero-gradient, fixed-gradient, fixed-value PINN BCs.
• Enable RTS for activation functions (generally torch::Model, using the RTTI in torch) so that we can include them in hyperparam tuning.
• Implement adaptive activation functions / gradient flow correction for the Adam solver.
• Investigate the residual loss w.r.t. OpenFOAM's operators.
• Implement the non-conservative version + cross-compare.
• Fix metadata (RESIDUAL_MSE and not GRAD_MSE).
• Add CI: build and visualization testing.
• Document.

If you would like to contribute to this repository, provide additional ideas or spark up a discussion, the most direct way to get in touch is to open an issue at:

https://github.com/AndreWeiner/mlfoam