Dyad is a new AI-enabled systems modeling tool which is being designed to unlock the workflows of the future that we have discussed in previous blog posts.
As a cloud-first, differentiable, and extensible system, Dyad is enabling a future where over-the-air updates of AI-integrated models can quickly and easily enhance product performance.
Model Development
The Dyad Language is at the heart of Dyad and provides a unified textual representation of the model, its graphics, and analyses that can be run on the model, from simulating over a timespan to autotuning a PID controller to discovering real dynamics via symbolic regression.
A purely textural representation of a model enables functional and accurate version control and traceability and also means that every plot and result can be faithfully and accurately reproduced. Tests can also be performed with semantic meaning, rather than simple regression tests on timeseries.
Dyad’s model development workflow is designed to achieve maximum composability between models while achieving maximal scalability. The Dyad Language is designed to work with the Julia JIT compiler stack and supports bi-directional editing of both graphics and the code. This textual formulation integrates with modern software engineering workflows via compatibility with Git and continuous integration and deployment (CI/CD). These features have been used at companies like Instron to design and maintain digital twins of sophisticated hydraulic crash equipment.
Generative AI for Model Translation
Modeling a new phenomenon can create a mental block regarding where to start. Engineering would typically have to spend hours scouring the literature in order to find how others have approached the problem and start by recreating similar models before venturing into the unknown. By using generative AI mixed with AI-based translation of models from the corpus of a multitude of modeling languages, Dyad can present natural starting points to the modeling process and overcome the activation energy of the starting writer's block.
Acausal Modeling Without Sacrificing Control Systems
Dyad uses an acausal formulation pioneered by tools such as Modelica. Acausal modeling allows the user to specify the high level physics of the system and allows an automated symbolic process to perform the derivation in order to arrive at the fundamental equations which determine the simulation.
Formulating acausal models can be a major boost to productivity and composability compared with causal system simulation tools, as shown by our Instron and NASA case studies.
However, these acausal tools often lack many of the features that causal control systems require, such as:
Auto-coding to C and embedded targets
Model linearization and causalized analysis points
Well-supported causal subsets of the language that allow control systems to integrate
Dyad bridges this gap between acausal modeling and control systems. Along with synchronous programming features and state machines, Dyad will be able to simulate a model all the way from regulated flight control systems to detailed physical models, and can generate binaries for a number of embedded targets. Additionally, Dyad leverages capabilities from the Julia language, such as novel state estimation methods to aid control system design, as done at Mitsubishi Electric Research Lab.
Through a mix of programmatic and LLM-enabled workflows, Modelica code can also be translated into Dyad code, including semantic processing.
Synchronous Programming
Dyad has the ability to build and represent discrete-time components which can mix with the continuous components. This allows complex controllers to be written using multiple clocks and state machines which are then interconnected with continuous plant models. The plant models can be automatically discretized to give fully discrete controllers which are then compatible with the code generation and deployment capabilities for deploying the controllers to hardware.
Scalable Compilers and Solver Integration
Dyad is a new generation of compilers that solves the issues of scaling to large-scale systems. Its infrastructure is not tied to a single solver but instead is able to make use of the entire Julia SciML stack with hundreds of different techniques, where some are optimized for small 8 ODE systems and others are GPU-parallel distributed and optimized for millions of equations spread over a supercomputer. Dyad’s compiler is able to leverage this large class of numerical infrastructure to re-specialize the approach based on different systems that are being modeled.
Dyad's infrastructure is structure-preserving, allowing it to retain code from arrays and loops. This enables it to integrate complex models, like partial differential equation discretizations used in computational fluid dynamics (CFD) and finite element models, with simpler system models. By combining this with specialized solvers, Dyad can generate efficient simulators that rival hand-tuned CFD codes, all within its composable modeling system.
While in isolation at the start of the project we do not expect a pure Dyad CFD model to be competitive with a code like Ansys Fluent that is hand-optimized for exactly that model form, the ability to seamlessly integrate and compose models will mean that combinations, like a spatial battery model cooled by a chiller, with a CFD model of the resulting airflow, can be greatly improved over situations which attempt to co-simulate independent simulation codes.
Model Refinement
Schematic of SciML-based automated model refinement from data. This showcases how recent advances such as differentiable programming and universal differential equations can be synchronized with streaming data sources in order to produce models which learn previously unknown physics on the fly, iteratively improving themselves at the pace of computing. This transforms the modeling workflow from being
The modeling and simulation tools of the Dyad stack are fully compatible with modern tooling for automatic differentiation (AD), enabling differentiable programming (dP) workflows. All forms of inverse problems, which includes the solving of problems like calibrating models to data and performing design optimizations, rely on a form of gradient-based optimization as the central calculation, and the calculation of the gradient is the core computation that takes the most time.
Dyad's automatic differentiation integration allows for the automatic construction of derivative functions that can provide gradients that are orders of magnitude faster and more accurate.
Git-based version control means that it's very easy for a team of modelers to iteratively work on and refine their Dyad model. Because there is a one-to-one mapping between GUI representations and the Dyad code, version diffs are representable graphically (with coloring and other tricks) able to highlight changes between model versions.
Improving Models with Real World Data
The integration of data into models and the resulting solution of inverse problems (neural network training, model calibration, etc.) can be a difficult computational task. Therefore it is necessary to track all of the experiments that were run and hyperparameters of the algorithms used (for example, results of different fitting choices, initial guesses, etc.) in order to get a full picture of the landscape. Dyad includes not only features for asynchronous execution of these long running jobs, i.e. the ability to spin up cloud runners to run these tasks in the backend, but also a complete logging system for investigating what types of analyses were previously performed and build comparisons between them.
Dyad is served through the JuliaHub platform, which provides connectors to many popular datalakes and datastores. After connecting with the appropriate data source, Dyad models can be deployed as live apps which can be retrained either at regular time intervals, or when a new batch of data enters the datastore, or externally via an API.
Neural Surrogates for Faster Training
Sometimes, a high fidelity source truth can be helpful to integrate into a system model. For example, precise simulators for fluid or structural properties can serve as a digital source of truth. The Dyad AI tooling for automated surrogate generation allows slimming of the twin of the component in a way that retains the required fidelity to integrate with the rest of the system without introducing heavy computational burden. This reduced computational burden has been leveraged to design control systems in both commercial and automotive HVAC systems, and even for solving large scale inverse problems when evaluating the erosion of public infrastructure such as bridges.
Model Analysis
One major differentiating factor of the Dyad system is its flexibility. This flexibility is not just within code and models but is also demonstrated in the analyses available in the GUI because Dyad includes a general system for hooking into the GUI for custom analyses. All of the internal analyses (controls, simulation, AI, etc.) are built to a documented API that exposes these model analyses to the user in the GUI.
This API is designed to be open and accessible to users as well, meaning that the user base can share libraries of models and also libraries of analyses. This means that specialized control-systems analyses, optimal experimental design techniques, etc. can be implemented as GUI extensions by external parties and shipped as part of our model library system in order to allow further growth of the Dyad platform.
But this also allows for company specific analyses, like the construction of specific plots or reports about models for regulatory deliverables, to be constructed and shared within a company and become a standard GUI button for a specific user group. This greatly grows the customizability of the GUI-based features and thus removes many of the limitations that are traditionally discussed about the legacy GUI-based system modeling tools.
Testing can also be done through analyses, so users can test properties of their components (does this converge, oscillate with a certain period, does this control the oscillation of the spring to some threshold, etcetera) directly.
Control Systems
Dyad includes a complete control-systems analysis suite which covers the traditional uses. This includes many features such as linear analyses (linearization, PID autotuning, etc.), robust control (H-infinity control), model-predictive control, and more. This functionality bridges the gap between what was traditionally covered by different offerings.
Dyad Success Stories
Instron: Dyad's underlying technology sped up Instron's "Catapult Light" simulation workflow by 500x, reducing simulation time from months to hours and enabling a lower-cost, high-performance product.
NASA Launch Services: Dyad's technology allowed NASA's RECURSAT launch services simulation to run 15,000x faster, cutting the time per run from 15 minutes to 58.2 milliseconds, which transformed a long, overnight analysis process into an interactive one.
Williams Racing: By creating a digital twin of a sensor, Dyad helped Williams Racing reduce prediction error by approximately 50% and evaluate models 4x faster compared to classic machine learning methods, all without the need for a physical sensor on the car.
Dyad: The Way Forward
Integrating AI into digital twins for industry requires care to maintain the safety and reliability of traditional engineering. Dyad is built to bring AI and SciML into the right places, helping engineers create more accurate digital models while keeping humans in the loop.
As adoption grows, we see models that continuously improve in the cloud using streaming data, with engineers validating updates and deploying them for predictive maintenance and over-the-air improvements. This shift could boost engineering productivity by an order of magnitude, enabling more efficient and reliable systems without compromising safety.
If you would like to explore Dyad's capabilities further for your specific use case, reach out to our product specialists.
