Software Defined Matter: An Open Substrate For Design And Optimization
Interview with Eamon Carrig – Emergent Matter
Eamon Carrig of Emergent Matter answers six questions on Software Defined Matter, a framework developed by the Austin based startup for differentiable design and manufacturing.
The interview covers the choice of signed distance fields, a single differentiable graph linking geometry, materials, physics and manufacturing constraints, and the .sdm file as an engineering record.
Can you introduce us to Emergent Matter, what the company does, and what your presentation will cover at CDFAM?
Emergent Matter is a new venture backed startup based in Austin, Texas. We exist to advance and commercialize state-of-the-art robotic systems.
Our founders worked together for the last 5 years at another company, where they ran over $175M in government R&D efforts surrounding cementitious and ceramic large scale additive manufacturing systems for extreme environments. In this presentation we will introduce the motivation for and praxis of a new framework we’ve developed. We call it Software Defined Matter, or SDM for short.
Writing a new design kernel, file format, AND approach to engineering from the ground up takes a particular mix of skills. Who is on the team, what backgrounds did they bring, and how did that combination lead to this approach?
The team includes physicists, engineers, and software developers. Most of what we’ve developed came from addressing pain points in our own careers designing, fabricating, and validating high performance machines. Above all else, we select for curiosity.
Signed distance fields are usually chosen for geometric flexibility, but here the emphasis is on differentiability. What does it take to make requirements, physics, materials and manufacturing constraints all differentiable alongside the geometry, and how does that connect with the other formats in an engineering chain and with the manufacturing process itself?
We chose signed distance fields because they make geometry an editable mathematical function rather than tying the design to a particular CAD surface or simulation mesh.
We’re building the core in a unified automatic-differentiation framework, connecting geometry, material models, coupled physics, and manufacturing constraints through a single computational graph. Each solver must return predictions and design sensitivities, using implicit or adjoint methods for expensive field solves.
The signed distance field representations describe geometry; SDM is the broader engineering record connecting that geometry to materials, process settings, requirements, and evaluation evidence.
Electromagnetic, thermal, and structural solvers can use different meshes while evaluating the same underlying design. Simulation, visualization, and manufacturing representations are derived from that shared source of truth rather than becoming independent versions of the part.
Speaking of data, the .sdm file is described as a single graph carrying parameters, geometry, materials, interfaces, objectives and constraints. How does data move through that graph during an optimization run, and how do external solvers, viewers and manufacturing tools read from and write back to the same source?
SDM is the structured engineering record of the design, not merely a geometry file with a history attached. Autodiff gives us both the sensitivities and the gradient, and backpropagating those sensitivities allows us to iteratively update our design towards an optimum.
External tools, viewers, and manufacturing tools can consume derived geometry and process data through adapters that translate the design into their native representations.
Design changes create new revisions rather than overwriting existing earlier designs & evidence. The source of truth is therefore the engineering record – not a requirement that every tool use the same mesh or file format, and that engineering record is the .sdm file.
Manufacturing limits look different depending on who is programming the machine. In house, with full control of the process, machines can be pushed. A contract manufacturer will tend toward the conservative end of the spectrum. How does the approach take both situations into account when bounding an optimization?
In-house, we can optimize process settings and part geometry together, exploring a wider operating range where testing supports it, and with a contract manufacturer we can work within their agreed process window, tolerances and restrictions. The method stays the same; what changes is which parameters the optimizer can adjust and which limits it must respect.
What do you hope attendees take away from the presentation at CDFAM, and what are you hoping to learn from the other presenters and attendees?
We hope to convince the audience that this approach is broadly useful and, by virtue of SDM, immediately usable.
Eamon Carrig presents Software Defined Matter at CDFAM Tokyo on October 8-9, where attendees can question the team directly and compare the approach with others on the program.
The event brings together those advancing computational design, AI and machine learning in design, engineering and architecture.
