Text-to-Spaceship: Accelerating Mission Development with AI at NASA
Interview with Ryan McClelland, AI Infusion Lead at NASA Goddard
Six questions with Ryan Mclelland, AI Infusion Lead, at NASA Goddard Space Flight Center ahead of his keynote presentation CD/DC, the next installment of the CDFAM Computational Design Symposium in Washington DC, July 15-16, 2026.
- When you first pitched ‘Text to Spaceship’ at CDFAM in 2024‘, it was an ambitious concept given the state of AI for engineering at the time. How have the available tools changed your approach to implementing the vision?
It has been wild to see Text-to-Spaceship go from an aspirational vision to a burgeoning reality so quickly; the consensus today is that ‘Text-to-X is happening now.’
Looking back, I don’t see big changes to the core concept, but the implementation approach has come increasingly into focus. A huge accelerating factor is the rise of Agentic software engineering tools like Claude Code and Codex. These tools enable non-developer engineers to create workflow automations much faster than previously expected.
While LLMs continue to improve, the ability of discipline experts to rapidly create deterministic or intelligent workflow automations is a significant uplift.
This advancement in Agentic software engineering both accelerates and provides a clear playbook for Agentic hardware engineering to follow.
Crucially, the implementation relies on a hybrid approach where the LLM handles the natural language requirements, but the computational design engine, often deterministic and physics-based process create designs, distinguishing it from the probabilistic “hallucination” prone models used for text and art.
- How are you now using AI in both formulating and executing the Text-to-Spaceship?
Frankly, looking back over the last few years of my work, it’s become difficult to disentangle which ideas originated with me or my team, and which came from AI.
I use these tools constantly, pushing at the edges of the ‘jagged frontier’ of capability. In formulation, AI is used for early ideation to explore very wide trade spaces and can even help us decide what to build, not just how, by referencing high-level directives like the “Moon to Mars architecture”.
This requires adopting a conversational approach with the LLM, thinking of it as a team member.
AI helps by giving me new ideas, but it also sometimes encodes incorrect consensus views.
For instance, when I show AI our Evolved Structures, it always recommends Additive Manufacturing, even though I know from direct trades that they should be, and almost always are, CNC machined.
Despite the limitations, the amount of work I can execute today on the Text-to-Spaceship vision is massively uplifted by AI tools, a benefit that is true for thousands of NASA employees’ daily work. Which cognitive tasks to keep and which to delegate to AI requires constant intention.
- Can you tell us about the software ecosystem you are connecting, including tools developed internally at NASA, third-party AI platforms, and software partners?
The grand vision is to build a multiplayer JARVIS for space mission development.
This ecosystem involves Text-to-X building blocks deployed in our secure cloud, connected and orchestrated by AI agents. We maintain a constant ‘horizon scan’ on the market, though the pace of development makes it impossible to evaluate every new product.
We handle our software needs in two main ways: Internal Development, building specialized tools unique to Goddard’s work, and Third-Party & Partners, purchasing tools that fit our needs and providing feedback to help shape their products. Examples of companies we are actively working with include Celedon, Synera, and Infinitform.
A significant hurdle slowing our progress is that many legacy engineering tools from the desktop days lack modern APIs, which can restrict us to using sluggish, CPU-reliant solvers. Similarly, the lack of APIs for industrial suppliers creates a friction point where agents are forced to navigate browsers just to fetch CAD files manually.
The current mainstream engineering ecosystem was not built with modern, Agentic API access in mind. To achieve true scale, we must move toward Agentic protocols with our vendors and adopt modern, cloud-based API-first CAE tools.
- Data is a recurring challenge across the industry, whether too much or too little. What role does simulation and synthetic data play in your work?
Simulation is essential to the feedback loops that drive design decisions. AI/ML can help here by both orchestrating complex simulations and accelerating them with surrogate models.
However, we are not focused on collecting massive amounts of data to create ‘one-shot’ AI design models or delving into synthetic data deeply yet. Instead, we are pursuing the Agentic Engineering approach where physics simulation is kept within the design harness.
I am not personally convinced that the approach of collecting massive amounts of data to create ‘one-shot’ AI design models will work for complex space hardware.
Instead, we are using the Agentic Engineering approach that has proven so successful in software. In this model, physics simulation is kept within the design harness since it is a verifiable process, rather than being hidden within the model’s weights and biases. While ‘world models’ hold promise, current efforts are primarily focused on humanoid robotics and have limited immediate application to the extreme environment of space. Perhaps the agentic harnesses to physics solvers are a step on the path to large engineering models useful for real-world design applications.
- Companies like Boston Dynamics are making significant progress in teaching autonomous systems to operate in complex environments. Do you foresee a future where the design, engineering, and manufacturing loop becomes similarly autonomous, driven by mission parameters and environmental feedback?
Absolutely! I foresee a fully autonomous loop. The key to realizing this is creating accurate, scalable, and fast physics simulations of those environments, since building physical hardware is relatively slow and expensive.
However, physical testing must still play a crucial role. We want to run ‘hardware-rich’ programs where physical test cycles are accelerated by faster design and manufacturing loops.
This hardware-rich approach is essential for reducing model uncertainty and demonstrating successful fabrication, ultimately closing the autonomous loop.
- Finally, what do you think we need to do to prepare the next generation of designers and engineers to tackle the unknowable challenges they will inevitably face?
I think we can look to the software engineering field for inspiration.
Just as the best professors are now training students to tackle the higher-level tasks of software architecture, the same shift must occur in hardware engineering. However, the Pareto front of what is possible in design is highly dependent on manufacturing capability.
Therefore, I see a major opportunity to increase focus on manufacturing innovation, which will expand design possibilities and accelerate the iteration of physical systems
Join us in Washington DC to connect in person with Ryan and other experts from industry, software development and academia exploring, and implementing the ‘jagged edge’ of AI and machine learning in design, engineering and architecture.
