JuliaHub Blog: Insights & Updates

The Strategic Connection Between JuliaHub, Dyad and the Julia Open Source Community

Written by Dr. Chris Rackauckas | Jul 03, 2025

Watch the original video: This blog post summarizes my presentation at the OSA Community event. You can find the original video here. (Note: JuliaSim is now known as Dyad.)

Introduction

The relationship between open source software and commercial products has been contentious for years. Most people assume that commercializing open source inevitably leads to community fragmentation or the tragedy of the commons. As VP of Modeling and Simulation at JuliaHub and a leading figure in the Julia scientific computing ecosystem, I want to share a different perspective.

At a recent OSA Community event, I presented what might be a surprising thesis: making the right software into products has actually expanded our open source community. The story of Julia, SciML, PumasAI, and Dyad (formerly JuliaSim) demonstrates how commercial success can directly fuel open source innovation rather than undermine it.

The Foundation: Julia and Scientific Computing

Julia's Promise: Speed Meets Simplicity

Julia emerged to solve the "two-language problem" that plagued scientific computing. Researchers would prototype in high-level languages like Python or MATLAB for ease of use, then rewrite performance-critical code in C or Fortran. Julia promised to eliminate this friction by being both expressive and fast.

 

The results have been impressive. Julia occupies the coveted bottom-left quadrant of programming language performance charts, combining Python's simplicity with C's speed. This foundation enabled us to develop SciML (Scientific Machine Learning), which has grown into an ecosystem of over 200 packages.

At the core of SciML is DifferentialEquations.jl, which has demonstrated remarkable performance advantages. Comprehensive benchmarks comparing ODE solver performance across languages show DifferentialEquations.jl consistently outperforming established libraries:

 

The numbers are striking: 50x faster than SciPy for differential equations, 50x faster than MATLAB, 100x faster than R's deSolve, and 20-100x faster than JAX and PyTorch for GPU parallelized ODEs. These benchmarks cover various problem types from simple non-stiff ODEs to complex stiff systems, demonstrating consistent performance advantages across different computational scenarios.

The Academic Foundation: MIT Julia Lab

 

The MIT Julia Lab, led by Alan Edelman, serves as the academic backbone of the Julia ecosystem. This is where Julia was originally created by Edelman along with Stefan Karpinski, Viral Shah, and Jeff Bezanson. The lab produces numerous influential packages and maintains a steady pipeline of PhD students who become major contributors to the open source ecosystem.

Faculty like Alan Edelman run major courses that train new Julia developers (his Linear Algebra course is particularly influential), while Chris Rackauckas serves as Co-PI developing the entire SciML organization. Current students include Avik Pal, who created Lux.jl for deep learning, and Bowen Zhu, who is working on Symbolics.jl, among many others making significant contributions.

Recent alumni have built foundational tools: Torkel Loman created Catalyst.jl for systems biology, Shashi Gowda developed Symbolics.jl, and Valentin Churavy built Julia's HPC and GPU stack. The list of contributors is extensive.

  • Alan Edelman (Principal Investigator): Runs major courses every year that train new students in Julia (Linear Algebra), writes many online training materials
  • Chris Rackauckas (Co-Principal Investigator and Research Affiliate): Lead developer of the Julia SciML organization and many of its packages, including DifferentialEquations.jl and ModelingToolkit.jl

Current Students and Postdocs:

  • Avik Pal (PhD Student): Created and maintains the Lux.jl deep learning system, maintains many of the SciML and deep learning packages
  • Bowen Zhu (PhD Student): Becoming a maintainer of Symbolics.jl
  • Nicholas Klugman (PhD Student): Creator of NeuralLyapunov.jl
  • Utkarsh (PhD Student): Major contributions to SciML packages, lead developer of DiffEqGPU.jl
  • Raye Kimmerer (Master's Student): Chair for JuliaCon, maintainer for Julia sparse linear algebra libraries and LinearSolve.jl
  • Songchen Tan (PhD Student): Creator of TaylorDiff.jl, getting involved in SciML packages
  • Vaibhav Dixit (Master's Student): Creator and maintainer of Optimization.jl and related SciML packages
  • Julian Samaroo (Research Software Engineer): Creator of Julia's AMD GPU support stack, maintainer of Dagger.jl

Recent Alumni:

  • Torkel Loman (Postdoctoral Associate): Creator of Catalyst.jl for systems biology, runs many community events and teaching workshops, maintainer of many SciML packages
  • Frank Schaefer (Postdoctoral Associate): Maintainer of many SciML packages including SciMLSensitivity.jl
  • Shashi Gowda (PhD Student): Creator of Symbolics.jl
  • Valentin Churavy (PhD Student): Creator of Julia's HPC and GPU stack, creator of Enzyme automatic differentiation library and KernelAbstractions.jl
  • Gaurav Arya (Undergraduate): Creator of StochasticAD.jl

However, there is a fundamental challenge with academic-centered open source development:

The Academic Sustainability Problem

Academic positions have natural time limits. PhD students graduate in 5-6 years, postdocs move on in 2-3 years, creating constant turnover of productive contributors. Securing funding requires writing lengthy grants with low acceptance rates, universities take substantial overhead, and funding agencies prioritize novel research over software maintenance.

Perhaps most problematically, the moment someone becomes highly productive with a library that gains widespread adoption, they graduate and leave. Success in creating useful software often coincides with completing academic requirements and moving on.

A Previous Commercialization Success: PumasAI

Identifying the Right Product-Market Fit

PumasAI emerged from earlier work at the University of Maryland Baltimore that collaborated with JuliaComputing (now JuliaHub) on nonlinear mixed effects modeling (NLME) for pharmacometrics—using mathematical models to determine optimal drug dosing for different patient populations. This work builds on the SciML differential equation solvers while adding specialized pharmaceutical tooling.

The decision to commercialize Pumas, the main product of PumasAI, rather than keep it open source was strategic. Pharmaceutical companies are heavy software users but rarely contribute code due to proprietary data and IP policies. They require FDA approval processes involving millions in investment and formal validation procedures. They need specialized features like regulatory report generation and compliance workflows.

Importantly, industry has established methods for paying for software but lacks mechanisms for reliable open source donations. This creates a structural advantage for commercialization. People within pharmaceutical companies may want to support open source communities but cannot provide "charity funding" through normal corporate channels. Corporate procurement processes aren't designed for charitable donations.

However, these same companies have well-established software acquisition channels. By creating a commercial product that interfaces with the open source ecosystem, internal champions can purchase software through normal procurement, with funding flowing back to open source development.

The Symbiotic Relationship

Rather than extracting value from the open source community, PumasAI has strengthened it by creating sustainable funding. When Big Pharma companies pay for PumasAI licenses and consulting, they indirectly fund core Julia infrastructure that benefits the entire scientific computing community.

This model works because pharmaceutical companies get the specialized tools and regulatory compliance they need, while their payments support the employment of key open source maintainers. The company employs numerous key maintainers of Julia packages, providing them with stable full-time positions to continue their open source work.

This creates a virtuous cycle: pharmaceutical companies receive specialized tools and regulatory compliance, funding continued development of open source foundations that enable even better future tools. Based on this model, PumasAI currently employs the following contributors to Julia packages

  • Dilum Aluthge: One of the key maintainers of the Julia package ecosystem repository and a major contributor to the full Julia CI/CD system that keeps the entire package ecosystem running smoothly.
  • Niklas Korsbo: Creator of Latexify.jl for LaTeX equation generation, with contributions to Catalyst.jl and many solid bug reports and minimal working examples (MWEs) that help improve package quality. He also creates training materials for the Julia community.
  • Michael Hatherly: Creator of the Documenter.jl documentation system for Julia and creator of many Julia packages for markdown support. Co-implemented the native Julia backend for technical publishing system Quarto (QuartoNotebookRunner.jl) together with Julius Krumbiegel.
  • Julius Krumbiegel: One of the top 3 contributors to Makie.jl (Julia's premier plotting ecosystem), the current maintainer of AlgebraOfGraphics, creator of Chain.jl for data processing pipelines, and DataFrameMacros.jl
  • Patrick Kofod Mogensen: Maintains Optim.jl (currently the most used nonlinear optimization package in Julia) and the JuliaNLSolvers organization with approximately 5 other packages.
  • Andreas Noack: Developer and maintainer of Julia's core Linear Algebra system and big contributor to statistics libraries.
  • Lucas Pereira: Maintains TopOpt.jl and contributes to various optimization-related packages.
  • Mohamed Tarek: Significant contributions to TopOpt.jl (topology optimization), AbstractDifferentiation.jl, and Nonconvex.jl, along with numerous helpful posts in Julia discourse helping the open source community.
  • David Widmann: Main contributor to and maintainer of Distributions.jl, one of the most used Julia packages. A steering council member for SciML and one of the top contributors to DifferentialEquations.jl (ranked #4 all-time contributor). He's also a top maintainer of Bayesian statistics packages in the Turing.jl ecosystem.

This employment model creates a virtuous cycle: pharmaceutical companies pay for specialized domain expertise and regulatory-compliant tools, which funds the continued development of the open source foundations that enable even better tools in the future. The result is a sustainable ecosystem where commercial success directly translates to open source advancement.

Moderna Therapeutics, a major pharmaceutical company, noted: "Pumas has emerged as our 'go-to' tool for most of our analyses in recent months... We are impressed by the quality and breadth of the experience of Pumas-AI scientists." This speaks to a key insight: these scientists are not only top experts in pharmacometrics and their respective fields, but also leading contributors to open source development. By creating a business model that allows the best people to continue doing open source work while solving real-world pharmaceutical problems, everyone benefits - industry gets cutting-edge tools from domain experts, and the broader scientific computing community gets sustained development of foundational open source packages.

The Next Level: Dyad (formerly JuliaSim)

What is Dyad?

Dyad represents the next evolution of this sustainable open source model, building on lessons learned from PumasAI's success. Built on the same SciML foundation, Dyad targets industrial modeling and simulation, competing with established tools like MATLAB/Simulink, Dymola, and Modelica.

This encompasses applications across aerospace and automotive engineering, building and monitoring wastewater treatment facilities, designing wind turbines, and other complex industrial engineering systems requiring sophisticated physics-based modeling and simulation. Before exploring how Dyad connects back to the open source community, it's important to understand what this product actually does and why it represents a significant advancement in industrial simulation.

Key Technical Innovations

Dyad introduces several breakthrough capabilities that distinguish it from traditional modeling tools:


Unified Language Architecture: Unlike traditional workflows requiring multiple tools and languages, everything from modeling to deployment happens in Julia. This eliminates friction from translating between different environments while maintaining consistency across the development pipeline. 

Scientific Machine Learning Integration: Dyad combines first-principles physics with data-driven machine learning approaches. This hybrid methodology produces more accurate and robust models than either pure physics-based or pure ML approaches alone.

Modern Cloud-Native Workflow: The platform embraces contemporary software development practices with cloud-native architecture, version control integration, and CI/CD-enabled development processes. This bridges the gap between traditional engineering workflows and modern software development practices.

Automated Digital Twin Generation: One of Dyad's most innovative features is automatically generating digital twins through a five-step process: create first-principle models, identify missing complexity requiring real-world data correction, gather data from deployed assets, use physics-informed neural networks to correct models with collected data, and deploy enhanced digital twins as FMUs or other universal formats.

Addressing Real Industrial Requirements

Dyad specifically targets the demanding requirements of industrial modeling:

  • Drag-and-drop model development for engineers familiar with graphical interfaces
  • Validated library of engineering models covering domains like HVAC, batteries, and multibody dynamics
  • Handling complex real-world scenarios including stiffness, differential algebraic equations (DAEs), events, contact modeling, and state machines
  • Strict code generation requirements meeting standards like DO-178, DO-330, and DAL-A for aerospace applications
  • Accessible SciML capabilities that don't require a machine learning background to leverage

The Workflow Revolution

Dyad's workflow spans three main phases:

  1. Model Building: Using both visual drag-and-drop interfaces and IDE-based coding in a unified environment
  2. Functionality Addition: Through JuliaSim Studio (VS Code integration) that allows seamless interaction between graphical models and code
  3. Deployment: Via JuliaHub/App for cloud-based simulation and analysis

This integrated approach means models exist as proper Julia packages with documentation, unit testing, and version control - treating engineering models with the same rigor as modern software development.

How Dyad Benefits from the Julia SciML Open Source Community

Performance and Scaling Advantages

Dyad's competitive advantages stem directly from its foundation on the Julia SciML open source ecosystem. Rather than building simulation capabilities from scratch, Dyad leverages over a decade of open source development and research, providing access to state-of-the-art numerical methods and performance optimizations that would be prohibitively expensive for any single company to develop independently.

The underlying SciML foundation provides significant performance advantages:

  • Automatic, inbuilt GPU acceleration on CUDA, AMD, Intel, and Apple GPUs through packages like CUDA.jl and AMDGPU.jl, built by the open source community,
  • advanced solvers from DifferentialEquations.jl that often outperform traditional Fortran implementations,
  • scalable cloud-native design built on Julia's parallelism capabilities,
  • scientific machine learning through NeuralPDE.jl and DataDrivenDiffEq.jl, that provide smoother and more accurate models than traditional ML approaches,
  • symbolic computation integration with Symbolics.jl for automatic code generation, and
  • automatic differentiation through Enzyme.jl and ForwardDiff.jl, that allow Dyad to automatically compute optimized gradients for parameter estimation and optimization tasks.

This foundation allows Dyad to offer capabilities that would take competing commercial tools years and millions of dollars to develop, while benefiting from continuous improvements made by hundreds of open source contributors worldwide.

How Dyad Improves the Open Source Ecosystem

Dyad's commercial success directly funds critical open source development through full-time employment of key Julia ecosystem contributors.

Julia Core Development: JuliaHub (Dyad's parent company) employs and has employed Julia's co-founders and core compiler developers:

  • Viral Shah (JuliaHub co-founder and CEO): Co-creator of Julia, continues to help with linear algebra and binaries
  • Jeff Bezanson (JuliaHub co-founder and CTO): Co-creator of Julia, lead developer of the Julia compiler
  • Matt Bauman (JuliaHub Director of Customer Success): Many contributions to Julia's Base library (arrays), moderator for Julia community forums
  • Avik Sengupta (JuliaHub VP of Engineering): Moderator for Julia community forums
  • Keno Fischer (JuliaHub co-founder and CTO): Major contributor to the Julia compiler
  • Stefan Karpinski (JuliaHub co-founder and CPO): Co-creator of Julia, continues to develop the Julia package manager
  • Gabriel Baraldi (JuliaHub Compiler Engineer): Major contributions to the Julia compiler

SciML and Scientific Computing: The company has funded maintainers of core SciML packages:

  • Alex Jones: Maintains MethodOfLines.jl and other SciML PDE tools
  • Shashi Gowda: Maintainer of ModelingToolkit.jl and other base SciML libraries
  • Yingbo Ma: Major contributions to the Julia compiler and SciML optimization
  • Dhairya Gandhi: Major contributor to Flux.jl deep learning library, major contributions to the Julia compiler
  • Fredrik Ekre: Creator and maintainer of the Ferrite FEM library
  • Sebastian Micluța-Câmpeanu: Major contributions to SciML and JuliaDynamics
  • Chris Rackauckas: Lead developer of the Julia SciML organization and many of its packages, including DifferentialEquations.jl and ModelingToolkit.jl

Julia Infrastructure and Community: Critical community infrastructure maintained by JuliaHub-funded developers:

  • Sebastian Pfitzner: Creator and maintainer Julia VS Code plugin
  • Morten Piibeleht: Maintainer of Documenter.jl and Julia's Documentation
  • Tim Besard: Creator and maintainer of the Julia GPU stack and CUDA.jl

The Feedback Loop: Commercial Needs Drive Open Source Innovation

Dyad's industrial requirements create a virtuous cycle that strengthens the open source ecosystem. For example, industrial-scale simulations stress-test SciML solvers, revealing performance bottlenecks that might not appear in academic use cases. Commercial customers demand reliability and edge-case handling, leading to more robust implementations.

  1. Performance Pressure: Industrial-scale simulations stress-test the SciML solvers, revealing performance bottlenecks and edge cases that might not appear in academic use cases. This drives optimization work that benefits all users.
  2. Robustness Requirements: Commercial customers demand reliability and edge-case handling, leading to more robust implementations in the underlying open source packages.
  3. Safety and Regulatory Standards: The safety requirements of highly regulated industries, such as aerospace and automotive, provide crucial feedback that improves the maintenance structures of the open source community. These industries demand rigorous development processes, comprehensive testing frameworks, and formal verification procedures that make the entire ecosystem more structured, stable, and focused on correctness. Standards like DO-178, DO-330, and DAL-A drive improvements in code quality, documentation, and testing that benefit all Julia users.
  4. Feature Development: Industrial workflows require capabilities like advanced visualization (Makie.jl), symbolic computation (Symbolics.jl), and GPU acceleration - all of which are developed as open source packages that the broader community can use.
  5. Integration Testing: Dyad's need to integrate multiple packages reveals compatibility issues and drives better package interoperability across the Julia ecosystem.
  6. Documentation and Usability: Commercial products require excellent documentation and user experience, improvements that flow back to the open source packages.

This model ensures that the open source foundations remain cutting-edge and well-maintained, funded by industrial applications while serving the broader scientific computing community.

High Level Summary of Benefits to JuliaHub from Connecting Dyad to the Open Source Community

What Makes This Work

  1. Choose the Right Products: Focus on specialized, domain-specific applications rather than core infrastructure
  2. Maintain Open Foundations: Keep the fundamental tools open source while commercializing the specialized applications
  3. Hire from the Community: Employ existing contributors rather than trying to build separate teams
  4. Strategic Positioning: Target users who benefit greatly from the software but cannot easily contribute back

Conclusion: Rethinking Open Source Sustainability

The Julia ecosystem demonstrates that strategic commercialization can strengthen rather than weaken open source communities. Through PumasAI and now Dyad, we have shown how identifying the right products to commercialize—those serving specialized needs of well-funded users who cannot easily contribute back—can create sustainable funding for broader open source infrastructure.

To conclude:

By making the right software into products, we have greatly expanded the open source community." This isn't about choosing between open source idealism and commercial pragmatism—it's about finding synergies that make both thrive.

The lesson for other open source ecosystems is clear: sustainability doesn't require abandoning open source principles. It requires strategic thinking about which layers of the stack can and should be commercialized to fund development of foundational tools that benefit everyone.

This model offers a blueprint for other scientific computing and technical communities seeking to build sustainable ecosystems that serve both academic researchers and industrial users while maintaining the collaborative spirit that makes open source powerful.