<p>GridFire is a C++ library designed to perform general nuclear network evolution. It is part of the larger SERiF project within the 4D-STAR collaboration. GridFire is primarily focused on modeling the most relevant burning stages for stellar evolution modeling. Currently, there is limited support for inverse reactions. Therefore, GridFire has a limited set of tools to evolves a fusing plasma in NSE; however, this is not the primary focus of the library and has therefor not had significant development. For those interested in modeling super nova, neutron star mergers, or other high-energy astrophysical phenomena, we <b>strongly</b> recomment using <ahref="https://bitbucket.org/jlippuner/skynet/src/master/">SkyNet</a>.</p>
<p>GridFire is architected to balance physical fidelity, computational efficiency, and extensibility when simulating complex nuclear reaction networks. Users begin by defining a composition, which is used to construct a full GraphEngine representation of the reaction network. A GraphNetwork uses <ahref="https://reaclib.jinaweb.org/index.php">JINA Reaclib</a> reaction rates (<ahref="https://iopscience.iop.org/article/10.1088/0067-0049/189/1/240">Cyburt et al., ApJS 189 (2010) 240.</a>) along with a dynamically constructed network topology. To manage the inherent stiffness and multiscale nature of these networks, GridFire employs a layered view strategy: partitioning algorithms isolate fast and slow processes, adaptive culling removes negligible reactions at runtime, and implicit solvers stably integrate the remaining stiff system. This modular pipeline allows researchers to tailor accuracy versus performance trade-offs, reuse common engine components, and extend screening or partitioning models without modifying core integration routines.</p>
<p>4D-STAR is funded by European Research Council (ERC) under the Horizon Europe programme (Synergy Grant agreement No. 101071505: 4D-STAR) Work for this project is funded by the European Union. Views and opinions expressed are however those of the author(s) only and do not necessarily reflect those of the European Union or the European Research Council.</p>
<p>By far the easiest way to install is with pip. This will install either pre-compiled wheels or, if your system has not had a wheel compiled for it, it will try to build locally (this may take <b>a long time</b>). The python bindings are just that and should maintain nearly the same speed as the C++ code. End users are strongly encourages to use the python module rather than the C++ code.</p>
<p><b>Note</b>: Currently macOS x86_64 does <b>not</b> have a precompiled wheel. Do to that platform being phased out it is likely that there will never be precompiled wheels or releases for it. </p>
<p><b>Note:</b> macOS wheels were targeted to macOS 12 Monterey and should work on any version more recent than that (at least as of August 2025). </p>
<p><b>Note:</b> Linux wheels were compiled using manylinux_2_28 and are expected to work on Debian 10+, Ubuntu 18.10+, Fedora 29+, or CentOS/RHEL 8+ </p>
</blockquote>
<blockquoteclass="doxtable">
<p><b>Note:</b> If your system does not have a prebuilt wheel the source distribution will download from pypi and try to build. This may simply not work if you do not have the correct system dependencies installed. If it fails the best bet is to try to build boost >= 1.83.0 from source and install (<ahref="https://www.boost.org/">https://www.boost.org/</a>) as that is the most common broken dependency. </p>
</blockquote>
<h3><aclass="anchor"id="autotoc_md6"></a>
source</h3>
<p>The user may also build the python bindings directly from source</p>
</div><!-- fragment --><blockquoteclass="doxtable">
<p><b>Note:</b> that if you do not have all system dependencies installed this will fail, the steps in further sections address these in more detail. </p>
</div><!-- fragment --><p>This will generate incremental builds whenever source code changes, and you run a python script automatically (note that since <code>meson setup</code> must run for each of these it does still take a few seconds to recompile regardless of how small a source code change you have made). It is <b>strongly</b> recommended that developers use this approach and end users <em>do not</em>.</p>
<p>The easiest way to build GridFire is using the <code>install.sh</code> or <code>install-tui.sh</code> scripts in the root directory. To use these scripts, simply run:</p>
</div><!-- fragment --><p> The regular installation script will select a standard "ideal" set of build options for you. If you want more control over the build options, you can use the <code>install-tui.sh</code> script, which will provide a text-based user interface to select the build options you want.</p>
<p>Generally, both are intended to be easy to use and will prompt you automatically to install any missing dependencies.</p>
<p><b>Note:</b> On Ubuntu 22.04 the user needs to install boost libraries manually as the versions in the Ubuntu repositories are too old. The installer automatically detects this and will instruct the user in how to do this. </p>
<p><b>Note:</b> Windows is not supported at this time and <em>there are no plans to support it in the future</em>. Windows users are encouraged to use WSL2 or a Linux VM. </p>
<p><b>Note:</b> If <code>install-tui.sh</code> is not able to find a usable version of boost it will provide directions to fetch, compile, and install a usable version. </p>
<p>GridFire ships with an installer (<code>install.sh</code>) which is intended to make the process of installation both easier and more repeatable.</p>
<p>Both scripts are intended to automate installation more or less completely. This includes dependency checking. In the event that a dependency cannot be found they try to install (after explicitly asking for user permission). If that does not work they will provide a clear message as to what went wrong.</p>
<p>The TUI mode provides easy modification of meson build system and compiler settings which can then be saved to a config file. This config file can then be loaded by either tui mode or cli mode (with the <code>--config</code>) flag meaning that build configurations can be made and reused. Note that this is <b>not</b> a deterministically reproducible build system as it does not interact with any system dependencies or settings, only meson and compiler settings.</p>
<p><b>Note:</b><code>install-tui.sh</code> is simply a script which calls <code>install.sh</code> with the <code>--tui</code> flag. You can get the exact same results by running <code>install.sh --tui</code>. </p>
<p><b>Note:</b> Call <code>install.sh</code> with the <code>--help</code> or <code>--h</code> flag to see command line options </p>
</blockquote>
<blockquoteclass="doxtable">
<p><b>Note:</b><code>clang</code> tends to compile GridFire much faster than <code>gcc</code> thus why I select it in the above asciinema recording. </p>
<p><b>Note:</b> Depending on the ubuntu version you have the libboost-all-dev libraries may be too old. If this is the case refer to the boost documentation for how to download and install a version <code>>=1.83.0</code></p>
<p><b>Note:</b> On recent versions of ubuntu python has switched to being externally managed by the system. We <b>strongly</b> recommend that if you install manually all python packages are installed inside some kind of virtual environment (e.g. <code>pyenv</code>, <code>conda</code>, <code>python-venv</code>, etc...). When using the installer script this is handled automatically using <code>python-venv</code>. </p>
<p>As noted above <code>clang</code> tends to compile GridFire much faster than <code>gcc</code>. If your system has both <code>clang</code> and <code>gcc</code> installed you may force meson to use clang via environmental variables</p>
<p>GridFire uses C++23 features and therefore only compilers and standard library implementations which support C++23 are supported. Generally we have found that <code>gcc >= 13.0.0</code> or <code>clang >= 16.0.0</code> work well.</p>
<p>GridFire is organized into a series of composable modules, each responsible for a specific aspect of nuclear reaction network modeling. The core components include:</p>
<li><b>Engine Module:</b> Core interfaces and implementations (e.g., <code>GraphEngine</code>) that evaluate reaction network rate equations and energy generation. Also implemented <code>Views</code> submodule.</li>
<li><b>Engine::Views Module:</b> Composable engine optimization and modification (e.g. <code>MultiscalePartitioningEngineView</code>) which can be used to make a problem more tractable or applicable.</li>
<li><b>Partition Module:</b> Implements partition functions (e.g., <code>GroundStatePartitionFunction</code>, <code>RauscherThielemannPartitionFunction</code> (<ahref="https://www.sciencedirect.com/science/article/pii/S0092640X00908349?via%3Dihub]">Rauscher & Thielemann, 2000</a>) to weight reaction rates based on nuclear properties.</li>
<li><b>Solver Module:</b> Defines numerical integration strategies (e.g., <code>DirectNetworkSolver</code>) for solving the stiff ODE systems arising from reaction networks.</li>
<li><b>Python Interface:</b> Exposes <em>almost</em> all C++ functionality to Python, allowing users to define compositions, configure engines, and run simulations directly from Python scripts.</li>
<p>Generally a user will start by selecting a base engine (currently we only offer <code>GraphEngine</code>), which constructs the full reaction network graph from a given composition. The user can then apply various engine views to adapt the network topology, such as partitioning fast and slow reactions, adaptively culling low-flow pathways, or priming the network with specific species. Finally, a numerical solver is selected to integrate the network over time, producing updated abundances and diagnostics.</p>
<p>GridFire is, at its core, based on a series of <code>Engines</code>. These are constructs which know how to report information on series of ODEs which need to be solved to evolver abundances. The important thing to understand about <code>Engines</code> is that they contain all the detailed physics GridFire uses. For example a <code>Solver</code> takes an <code>Engine</code> but does not compute physics itself. Rather, it asks the <code>Engine</code> for stuff like the jacobian matrix, stoichiometry, nuclear energy generation rate, and change in abundance with time.</p>
<p>Refer to the API documentation for the exact interface which an <code>Engine</code> must implement to be compatible with GridFire solvers.</p>
<p>Currently, we only implement <code>GraphEngine</code> which is intended to be a very general and adaptable <code>Engine</code>.</p>
<p>In GridFire the <code>GraphEngine</code> will generally be the most fundamental building block of a nuclear network. A <code>GraphEngine</code> represents a directional hypergraph connecting some set of atomic species through reactions listed in the <ahref="https://reaclib.jinaweb.org/index.php">JINA Reaclib database</a>.</p>
<p><code>GraphEngine</code>s are constructed from a seed composition of species from which they recursively expand their topology outward, following known reaction pathways and adding new species to the tracked list as they expand.</p>
<li><code>BuildDepthType</code> (<code>Full</code>, <code>Shallow</code>, <code>SecondOrder</code>, etc...): controls number of recursions used to construct the network topology. Can either be a member of the <code>NetworkBuildDepth</code> enum or an integer.</li>
<li><code>partition::PartitionFunction</code>: Partition function used when evaluating detailed balance for inverse rates.</li>
<tdclass="markdownTableBodyNone"><code>RauscherThielemannPartitionFunction</code></td><tdclass="markdownTableBodyNone">"RauscherThielemann" </td><tdclass="markdownTableBodyNone">Interpolates normalized g-factors per Rauscher & Thielemann. </td></tr>
<trclass="markdownTableRowOdd">
<tdclass="markdownTableBodyNone"><code>CompositePartitionFunction</code></td><tdclass="markdownTableBodyNone">"Composite" </td><tdclass="markdownTableBodyNone">Combines multiple partition functions for situations where different partitions functions are used for different domains </td></tr>
<p>One of the primary tasks any engine must accomplish is to report the jacobian matrix of the system to the solver. <code>GraphEngine</code> uses <code>CppAD</code>, a C++ auto differentiation library, to generate analytic jacobian matrices very efficiently.</p>
<p>All reactions in JINA Reaclib which only include reactants iron and lighter were downloaded on June 17th, 2025 where the most recent documented change on the JINA Reaclib site was on June 24th, 2021.</p>
<p>All of these reactions have been compiled into a header file which is then statically compiled into the gridfire binaries (specifically into lib_reaction_reaclib.cpp.o). This does increase the binary size by a few MB; however, the benefit is faster load times and more importantly no need for end users to manage resource files.</p>
<p>If a developer wants to add new reaclib reactions we include a script at <code>utils/reaclib/format.py</code> which can ingest a reaclib data file and produce the needed header file. More details on this process are included in <code>utils/reaclib/readme.md</code></p>
<p>The GridFire engine supports multiple engine view strategies to adapt or restrict network topology. Generally when extending GridFire the approach is likely to be one of adding new <code>EngineViews</code>.</p>
<thclass="markdownTableHeadNone">View Name </th><thclass="markdownTableHeadNone">Purpose </th><thclass="markdownTableHeadNone">Algorithm / Reference </th><thclass="markdownTableHeadNone">When to Use </th></tr>
<trclass="markdownTableRowOdd">
<tdclass="markdownTableBodyNone">AdaptiveEngineView </td><tdclass="markdownTableBodyNone">Dynamically culls low-flow species and reactions during runtime </td><tdclass="markdownTableBodyNone">Iterative flux thresholding to remove reactions below a flow threshold </td><tdclass="markdownTableBodyNone">Large networks to reduce computational cost </td></tr>
<trclass="markdownTableRowEven">
<tdclass="markdownTableBodyNone">DefinedEngineView </td><tdclass="markdownTableBodyNone">Restricts the network to a user-specified subset of species and reactions </td><tdclass="markdownTableBodyNone">Static network masking based on user-provided species/reaction lists </td><tdclass="markdownTableBodyNone">Targeted pathway studies or code-to-code comparisons </td></tr>
<tdclass="markdownTableBodyNone">FileDefinedEngineView </td><tdclass="markdownTableBodyNone">Load a defined engine view from a file using some parser </td><tdclass="markdownTableBodyNone">Same as DefinedEngineView but loads from a file </td><tdclass="markdownTableBodyNone">Same as DefinedEngineView </td></tr>
<tdclass="markdownTableBodyNone">MultiscalePartitioningEngineView </td><tdclass="markdownTableBodyNone">Partitions the network into fast and slow subsets based on reaction timescales </td><tdclass="markdownTableBodyNone">Network partitioning following Hix & Thielemann Silicon Burning I & II (DOI:10.1086/177016,10.1086/306692) </td><tdclass="markdownTableBodyNone">Stiff, multi-scale networks requiring tailored integration </td></tr>
<trclass="markdownTableRowOdd">
<tdclass="markdownTableBodyNone">NetworkPrimingEngineView </td><tdclass="markdownTableBodyNone">Primes the network with an initial species or set of species for ignition studies </td><tdclass="markdownTableBodyNone">Single-species ignition and network priming </td><tdclass="markdownTableBodyNone">Investigations of ignition triggers or initial seed sensitivities </td></tr>
<p>These engine views implement the common Engine interface and may be composed in any order to build complex network pipelines. New view types can be added by deriving from the <code>EngineView</code> base class, and linked into the composition chain without modifying core engine code.</p>
<p>There are certain functions for which it is expected that a call to an engine view will propagate the result down the chain of engine views, eventually reaching the base engine (e.g. <code>DynamicEngine::update</code>). We do not strongly enforce this as it is not hard to contrive a situation where that is not the mose useful behavior; however, we do strongly encourage developers to think carefully about passing along calls to base engine methods when implementing new views.</p>
<p>GridFire defines a flexible solver architecture through the <code>networkfire::solver::NetworkSolverStrategy</code> interface, enabling multiple ODE integration algorithms to be used interchangeably with any engine that implements the <code>Engine</code> or <code>DynamicEngine</code> contract.</p>
<p>All GridFire solvers implement the abstract strategy templated by <code>NetworkSolverStrategy</code> which enforces only that there is some <code>evaluate</code> method with the following signature</p>
</div><!-- fragment --><p> Which is intended to integrate some network over some time and returns updated abundances, temperature, density, and diagnostics.</p>
<p>GridFire solvers use a unified input and output type for their public interface (though as developers will quickly learn, internally these are immediately broken down into simpler data structures). All solvers expect a <code>NetIn</code> struct for the input type to the <code>evaluate</code> method and return a <code>NetOut</code> struct.</p>
<p>><b>Note:</b> It is often useful to set <code>NetIn::dt0</code> to something <em>very</em> small and >let an iterative time stepper push the timestep up. Often for main sequence >burning I use ~1e-12 for dt0</p>
<p>><b>Note:</b> The composition must be a <code>fourdst::composition::Composition</code>>object. This is made available through the <code>foursdt</code> library and the ><code>fourdst/composition/Composition.h</code> header. <code>fourdst</code> is installed >automatically with GridFire</p>
<p>><b>Note:</b> In Python composition comes from <code>fourdst.composition.Composition</code>>and similarly is installed automatically when building GridFire python >bindings.</p>
<h4><aclass="anchor"id="autotoc_md40"></a>
NetOut</h4>
<p>A <code>NetOut</code> struct contains</p><ul>
<li>The final composition after evolving to <code>tMax</code> (<code>NetOut::composition</code>)</li>
<li>The number of steps the solver took to evolve to <code>tmax</code> (<code>NetOut::num_steps</code>)</li>
<li>The final energy generated by the network while evolving to <code>tMax</code> (<code>NetOut::energy</code>)</li>
</ul>
<p>><b>Note:</b> Currently <code>GraphEngine</code> only considers energy due to nuclear mass >defect and not neutrino loss.</p>
<li><b>Integrator:</b> Implicit Rosenbrock4 scheme (order 4) via <code>Boost.Odeint</code>’s <code>rosenbrock4<double></code>, optimized for stiff reaction networks with adaptive step size control using configurable absolute and relative tolerances.</li>
<li><b>Jacobian Assembly:</b> Asks the base engine for the Jacobian Matrix</li>
<li><b>Linear Algebra:</b> Utilizes <code>Boost.uBLAS</code> for state vectors and dense Jacobian matrices, with sparse access patterns supported via coordinate lists of nonzero entries.</li>
<li><b>Error Control and Logging:</b> Absolute and relative tolerance parameters (<code>absTol</code>, <code>relTol</code>) are read from configuration; Quill loggers, which run in a separate non blocking thread, capture integration diagnostics and step statistics.</li>
<li><b>Initialization:</b> Convert input temperature to T9 units, retrieve tolerances, and initialize state vector <code>Y</code> from equilibrated composition.</li>
<li><b>Integrator Setup:</b> Construct the controlled Rosenbrock4 stepper and bind <code>RHSManager</code> and <code>JacobianFunctor</code>.</li>
<li><b>Adaptive Integration Loop:</b><ul>
<li>Perform <code>integrate_adaptive</code> advancing until <code>tMax</code>, catching any <code>StaleEngineTrigger</code> to repartition the network and update composition.</li>
<li>On each substep, observe states and log via <code>RHSManager::observe</code>.</li>
</ul>
</li>
<li><b>Finalization:</b> Assemble final mass fractions, compute accumulated energy, and populate <code>NetOut</code> with updated composition and diagnostics.</li>
<li><b>Operator Splitting Solvers:</b> Strategies to decouple thermodynamics, screening, and reaction substeps for performance on stiff, multiscale networks.</li>
<li><b>Callback observer support:</b> Currently we use an observer built into our <code>RHSManager</code> (<code>RHSManager::observe</code>); however, we intend to include support for custom, user defined, observer method.</li>
<p>These strategies can be developed by inheriting from <code>NetworkSolverStrategy</code> and registering against the same engine types without modifying existing engine code.</p>
<p>Through the Python bindings, users can subclass engine view classes directly in Python, override methods like <code>evaluate</code> or <code>generateStoichiometryMatrix</code>, and pass instances back into C++ solvers. This enables rapid prototyping of custom view strategies without touching C++ sources.</p>
<divclass="ttc"id="aclassgridfire_1_1_graph_engine_html"><divclass="ttname"><ahref="classgridfire_1_1_graph_engine.html">gridfire::GraphEngine</a></div><divclass="ttdoc">A reaction network engine that uses a graph-based representation.</div><divclass="ttdef"><b>Definition</b> engine_graph.h:97</div></div>
<divclass="ttc"id="aclassgridfire_1_1_adaptive_engine_view_html"><divclass="ttname"><ahref="classgridfire_1_1_adaptive_engine_view.html">gridfire::AdaptiveEngineView</a></div><divclass="ttdoc">An engine view that dynamically adapts the reaction network based on runtime conditions.</div><divclass="ttdef"><b>Definition</b> engine_adaptive.h:49</div></div>
<p>A representative workflow often composes multiple engine views to balance accuracy, stability, and performance when integrating stiff nuclear networks:</p>
<divclass="fragment"><divclass="line"><spanclass="preprocessor">#include "<aclass="code"href="engine_8h.html">gridfire/engine/engine.h</a>"</span><spanclass="comment">// Unified header for real usage</span></div>
<divclass="ttc"id="aclassgridfire_1_1_multiscale_partitioning_engine_view_html"><divclass="ttname"><ahref="classgridfire_1_1_multiscale_partitioning_engine_view.html">gridfire::MultiscalePartitioningEngineView</a></div><divclass="ttdoc">An engine view that partitions the reaction network into multiple groups based on timescales.</div><divclass="ttdef"><b>Definition</b> engine_multiscale.h:180</div></div>
<divclass="ttc"id="aengine_8h_html"><divclass="ttname"><ahref="engine_8h.html">engine.h</a></div><divclass="ttdoc">Core header for the GridFire reaction network engine module.</div></div>
<li><b>GraphEngine</b> constructs the full reaction network, capturing all species and reactions.</li>
<li><b>MultiscalePartitioningEngineView</b> segregates reactions by characteristic timescales (Hix & Thielemann), reducing the effective stiffness by treating fast processes separately.</li>
<li><b>AdaptiveEngineView</b> prunes low-flux species/reactions at runtime, decreasing dimensionality and improving computational efficiency.</li>
<li><b>DirectNetworkSolver</b> employs an implicit Rosenbrock method to stably integrate the remaining stiff system with adaptive step control.</li>
</ul>
<p>This layered approach enhances stability for stiff networks while maintaining accuracy and performance.</p>
<p>Custom callback functions can be registered with any solver. Because it might make sense for each solver to provide different context to the callback function, you should use the struct <code><aclass="el"href="namespacegridfire_1_1solver.html">gridfire::solver</a>::<SolverName>::TimestepContext</code> as the argument type for the callback function. This struct contains all the information provided by that solver to the callback function.</p>
<divclass="fragment"><divclass="line"><spanclass="preprocessor">#include "<aclass="code"href="engine_8h.html">gridfire/engine/engine.h</a>"</span><spanclass="comment">// Unified header for real usage</span></div>
<divclass="line"><spanclass="preprocessor">#include "<aclass="code"href="solver_8h.html">gridfire/solver/solver.h</a>"</span><spanclass="comment">// Unified header for solvers</span></div>
</div><!-- fragment --><p>><b>Note:</b> A fully detailed list of all available information in the TimestepContext struct is available in the API documentation.</p>
<p>><b>Note:</b> The order of species in the boost state vector (<code>ctx.state</code>) is <b>not guaranteed</b> to be any particular order run over run. Therefore, in order to reliably extract </p><blockquoteclass="doxtable">
<p>values from it, you <b>must</b> use the <code>getSpeciesIndex</code> method of the engine to get the index of the species you are interested in (these will always be in the same order). </p>
<p>Since each solver may provide different context to the callback function, and it may be frustrating to refer to the documentation every time, we also enforce that all solvers must implement a <code>descripe_callback_context</code> method which returns a vector of tuples<string, string> where the first element is the name of the field and the second is its datatype. It is on the developer to ensure that this information is accurate.</p>
<p>The python bindings intentionally look <b>very</b> similar to the C++ code. Generally all examples can be adapted to python by replacing includes of paths with imports of modules such that</p>
<p>This example implements the same logic as the above C++ example </p><divclass="fragment"><divclass="line"><spanclass="keyword">from</span> gridfire.engine <spanclass="keyword">import</span> GraphEngine, MultiscalePartitioningEngineView, AdaptiveEngineView</div>
<divclass="line">print(f<spanclass="stringliteral">"Initial H-1 mass fraction {comp.getMassFraction("</span>H-1<spanclass="stringliteral">")}"</span>)</div>
<divclass="line">print(f<spanclass="stringliteral">"Final H-1 mass fraction {results.composition.getMassFraction("</span>H-1<spanclass="stringliteral">")}"</span>)</div>
<p>Just like in C++, python users can register callbacks to be called at the end of each successful timestep. Note that these may slow down code significantly as the interpreter needs to jump up into the slower python code therefore these should likely only be used for debugging purposes.</p>
<p>The syntax for registration is very similar to C++ </p><divclass="fragment"><divclass="line"><spanclass="keyword">from</span> gridfire.engine <spanclass="keyword">import</span> GraphEngine, MultiscalePartitioningEngineView, AdaptiveEngineView</div>
<divclass="line">print(f<spanclass="stringliteral">"Initial H-1 mass fraction {comp.getMassFraction("</span>H-1<spanclass="stringliteral">")}"</span>)</div>
<divclass="line">print(f<spanclass="stringliteral">"Final H-1 mass fraction {results.composition.getMassFraction("</span>H-1<spanclass="stringliteral">")}"</span>)</div>
<divid="nav-path"class="navpath"><!-- id is needed for treeview function! -->
<ul>
<liclass="footer">Generated by <ahref="https://www.doxygen.org/index.html"><imgclass="footer"src="doxygen.svg"width="104"height="31"alt="doxygen"/></a> 1.13.2 </li>
