The AI knowledge base for open source

OpenGeoSys (OGS) is a scientific open-source project for numerical simulation of thermo-hydro-mechanical-chemical (THMC) processes in porous and fractured media. It is implemented in modern C++ (C++23) with an object-oriented design focused on the numerical solution of coupled multi-field (multi-physics) problems. The project supports parallel computing through both MPI and OpenMP. Application domains include CO2 sequestration, geothermal energy, water resources management, hydrology, and nuclear waste disposal. The codebase is organised into layered libraries: foundational utilities (BaseLib, MathLib), numerical methods (NumLib), mesh and geometry handling (MeshLib, GeoLib), a material property library (MaterialLib/MPL), and a large process library (ProcessLib) containing over 20 coupled-process implementations. Each process follows a consistent pattern of Process, ProcessData, LocalAssembler, and factory classes. The main application entry point is a command-line simulator that reads XML project files (.prj) describing geometry, mesh, material properties, boundary conditions, and solver settings. A Qt-based Data Explorer provides visualisation capabilities. The Material Property Library (MPL) offers a flexible, component-phase-medium hierarchy for defining constitutive relationships. NumLib provides finite element shape functions, ODE solvers, Newton-Raphson iteration, time-stepping algorithms, and degree-of-freedom management. ChemistryLib integrates reactive transport via PHREEQC. The build system uses CMake 3.31+ with presets and the project is tested through both unit tests (Google Test) and integration tests driven by benchmark .prj files.

GuideLLM is an open-source platform for evaluating language model performance under realistic, production-like workloads. It was created by Neural Magic (now part of the vLLM project) to give ML and infrastructure teams a consistent, rigorous framework for measuring how LLM inference servers behave under load. The platform simulates end-to-end interactions with OpenAI-compatible and vLLM-native servers, generating traffic patterns that reflect real application usage rather than synthetic micro-benchmarks. GuideLLM captures comprehensive latency and throughput metrics — including Time to First Token (TTFT), Inter-Token Latency (ITL), and full output-token distributions — that are essential for SLO-driven capacity planning. It supports a wide range of scheduling strategies including synchronous, concurrent, constant-rate, Poisson-distributed, and automatic sweep modes to explore an inference server's operating envelope. Data inputs can come from HuggingFace datasets, local files, or synthetic generators, and the system supports text, image, audio, and video modalities through extensible preprocessing pipelines. A multi-process scheduler with pluggable worker groups enables high-throughput load generation well beyond what simple async scripts or single-threaded tools can achieve. Benchmark results are aggregated into structured reports with per-request statistics and full metric distributions, then exported in JSON, YAML, CSV, and interactive HTML formats. Pre-built scenarios (chat, RAG) bundle sensible defaults for common evaluation patterns, while a flexible CLI and Python API allow deep customization for specialized workloads. Configuration is managed through Pydantic-based settings classes that read from environment variables, making container-based deployments straightforward. A Sanic-powered mock server lets teams develop and validate benchmarking workflows without requiring a live model deployment. GuideLLM is designed to be extensible: new backends, data preprocessors, scheduling strategies, and output formats can be registered through a shared registry/mixin pattern. The project follows standard Python packaging conventions with uv and tox for dependency management and test orchestration, and includes unit, integration, and end-to-end test suites.

OpenMC is an open-source Monte Carlo particle transport code designed for simulating nuclear systems such as fission reactors, fusion devices, and radiation shielding problems. It was originally developed at MIT's Computational Reactor Physics Group and is now maintained by a broad community of contributors worldwide. The code performs continuous-energy and multigroup neutron and photon transport using the Monte Carlo method, where particles are tracked stochastically through complex geometries. OpenMC supports constructive solid geometry (CSG) using surfaces, regions, cells, universes, and lattices, as well as CAD-based geometry via DAGMC integration. The codebase is a hybrid C++17/Python architecture: the computationally intensive transport kernel lives in C++, while the Python layer provides a rich API for building models, processing output, and controlling simulations. A C API (capi.h) bridges the two layers, exposed to Python via ctypes bindings in the openmc.lib package. OpenMC supports both k-eigenvalue criticality calculations and fixed-source simulations, and also implements a deterministic random ray solver as an alternative to Monte Carlo transport. The depletion module (openmc.deplete) couples neutron transport to isotopic burnup equations, enabling fuel cycle analysis using various operator-splitting integrators. The tally system is highly flexible, supporting dozens of filter types (spatial, energy, reaction, etc.) and score types for extracting reaction rates and other physical quantities. Nuclear data is stored in HDF5 format and accessed through the openmc.data module, which also handles ENDF file parsing, ACE format processing, and windowed multipole resonance treatment. Simulation output is written to HDF5 statepoint files, which can be read back through openmc.StatePoint for post-processing tally results and restarting runs. OpenMC leverages OpenMP for shared-memory parallelism and MPI for distributed-memory parallel simulation. The Python model API is organized around a Model class that bundles Geometry, Materials, Settings, Tallies, and Plots, and can export to XML for the C++ engine to consume. OpenMC is widely used in academia and national laboratories for reactor physics research, nuclear data validation, and shielding analysis.

The Weather Research and Forecasting (WRF) Model is an open-source, community-supported mesoscale numerical weather prediction (NWP) system developed primarily at the National Center for Atmospheric Research (NCAR). Version 4.7.1 of the codebase supports both operational forecasting and atmospheric research applications across a wide range of spatial and temporal scales. The model's primary dynamical core is the Advanced Research WRF (ARW), which implements a fully compressible, Euler nonhydrostatic equation set on an Arakawa C-grid with terrain-following hybrid sigma-pressure vertical coordinates. It supports Runge-Kutta 2nd- and 3rd-order time-integration schemes with acoustic mode splitting for numerical stability. WRF features an extensive suite of interchangeable physical parameterization packages spanning microphysics, planetary boundary layer, surface layer, land surface, cumulus convection, shallow convection, radiation, and urban physics. An active data Registry system drives compile-time code generation for model state variables, I/O streams, nesting operations, and namelist configuration, making it straightforward to add new fields or schemes. The framework supports multi-level parallelism combining distributed-memory MPI (via the RSL_LITE communication layer) and shared-memory OpenMP threading. Flexible I/O backends are provided including NetCDF, parallel NetCDF, HDF5, GRIB1/2, and raw binary formats, with asynchronous quilting I/O. Two-way and one-way domain nesting, including moving nests for tropical cyclone tracking, are supported with configurable integer grid ratios and feedback. WRFDA, the WRF Data Assimilation system, implements 3DVAR, 4DVAR, hybrid ensemble-variational (3DEnVar/4DEnVar), and ensemble Kalman filter (ETKF) methods for ingesting dozens of conventional and satellite observation types. WRF-Chem couples an extensive atmospheric chemistry and aerosol module directly within the model time-stepping loop. WRF-Hydro provides an optional terrestrial hydrological routing layer, coupling the land surface model to channel, overland, and subsurface flow simulations. The WRFPLUS package adds tangent-linear and adjoint models for 4DVAR data assimilation and sensitivity studies. Together these subsystems make WRF one of the most widely used atmospheric modeling frameworks in the world, serving research institutions, operational centers, and university programs globally.