ModelingToolkit.jl is a symbolic modeling language for Julia built in Julia. Its goal is very different from Sympy: it was made to support symbolic-numerics, the combination of symbolic computing with numerical methods to allow for extreme performance computing that would not be possible without modifying the model. Because of this, ModelingToolkit.jl excels in many areas due to purposeful design decisions:
- Performance: ModelingToolkit.jl is built in Julia, whereas SymPy was built in Python. Thus the performance bar for ModelingToolkit.jl is much higher. ModelingToolkit.jl started because SymPy was far too slow and SymEngine was far too inflexible for the projects they were doing. Performance is key to ModelingToolkit.jl. If you find any performance issues, please file an issue.
lambdifyis "fine" for some people, but if you're building a super fast MPI-enabled Julia/C/Fortran simulation code, having a function that hits the Python interpreter is less than optimal. By default,
build_functionbuilds fast JIT-compiled functions due to being in Julia. However, it has support for things like static arrays, non-allocating functions via mutation, fast functions on sparse matrices and arrays of arrays, etc.: all core details of doing high performance computing.
- Parallelism: ModelingToolkit.jl has pervasive parallelism. The symbolic simplification via SymbolicUtils.jl has built-in parallelism, ModelingToolkit.jl builds functions that parallelizes across threads and multiprocesses across clusters, and it has dynamic scheduling through tools like Dagger.jl. ModelingToolkit.jl is compatible with GPU libraries like CUDA.jl.
- Scientific Machine Learning (SciML): ModelingToolkit.jl is made to synergize with the high performance Julia SciML ecosystem in many ways. At a base level, all expressions and built functions are compatible with automatic differentiation like ForwardDiff.jl and Zygote.jl, meaning that it can be used in and with neural networks. Tools like DataDrivenDiffEq.jl can reconstruct symbolic expressions from neural networks and data while NeuralNetDiffEq.jl can automatically solve partial differential equations from symbolic descriptions using physics-informed neural networks.
- Primitives for high-performance numerics. Features like
ODESystemcan be used to easily generate automatically parallelized ODE solver code with sparse Jacobians and all of the pieces required to get the most optimal solves. Support for differential-algebraic equations, chemical reaction networks, and generation of code for nonlinear optimization tools makes ModelingToolkit.jl a tool for, well, building, generating, and analyzing models.
- Deep integration with the Julia ecosystem: ModelingToolkit.jl's integration with neural networks is not the only thing that's deep. ModelingToolkit.jl is built with the same philosophy as other SciML packages, eschewing "monorepos" for a distributed development approach that ties together the work of many developers. The differentiation parts utilize tools from automatic differentiation libraries, all linear algebra functionality comes from tracing Julia Base itself, symbolic rewriting (simplification and substitution) comes from SymbolicUtils.jl, parallelism comes from Julia Base libraries and Dagger.jl, and etc. The list keeps going. All told, by design ModelingToolkit.jl's development moves fast because it's effectively using the work of hundreds of Julia developers, allowing it to grow fast.