Research
The hard parts, stated plainly.
Taking an idea to a physical object with AI is not solved. This is where Terranoux is pointed, what is still open, and what is being worked on now.
Technical direction
Six lines of work.
- 01
A connected engineering representation
Requirements, decisions, geometry, models, simulations, experiments, components, cost and manufacturing held as one linked structure, so a change in one place can be traced to everything it affects.
- 02
Reasoning under physical constraints
Generating options is cheap. Rejecting the ones that violate physics, tolerances, materials or cost, and explaining why, is the work.
- 03
Simulation with known validity
Treating simulation as a tool with a stated range of validity, not an oracle. Every result should say which model produced it and where that model is known to fail.
- 04
Experiments that close the loop
Choosing the smallest physical test that resolves the largest uncertainty, running it inside an envelope, and feeding raw evidence back into the design.
- 05
Manufacturing knowledge as data
Process limits, machine capabilities, supplier options and cost curves represented well enough that a design can be checked against them while it is still a sketch.
- 06
Governed physical action
Purchases, equipment operation and orders proposed by Terranoux and evaluated by LucidRail, with every grant and refusal recorded.
Initial validation domain
Mechanical components first.
The ambition covers the whole lifecycle and many disciplines. The first proof should be narrow enough to measure. Single mechanical parts are where predictions can be checked against reality quickly and cheaply.
Validation path for one part
- 01
Requirements
loads, interfaces, envelope, cost
- 02
CAD
dimensioned and toleranced geometry
- 03
Analysis
FEA with stated assumptions
- 04
Manufacturability
process, tolerances, cost check
- 05
Make
machined, bent or printed
- 06
Inspect
dimensions against drawing
- 07
Load test
strain and failure against prediction
- 08
Compare
predicted vs measured, published
Steps 5 to 8 happen in the physical world, through partners and existing equipment. Success means predictions that hold within stated uncertainty, parts that fit and can be made at the estimated cost, and failures that are reported, not hidden.
Limitations and open problems
What we do not know how to do yet.
- Simulation fidelity
- Models drift from reality at contacts, interfaces, wear and long time scales. Knowing when to stop trusting a simulation is unsolved in general.
- Geometry for real parts
- Producing CAD that is manufacturable, dimensioned and toleranced, not just plausible-looking shapes.
- Sparse material data
- Properties depend on process, orientation and history. Published values are often not what a real part will see.
- Tacit manufacturing knowledge
- Much of what makes a part easy or hard to make lives in people, not documents.
- Verifying AI decisions
- An engineering decision needs a reason a human can check. Plausible but wrong is worse than no answer.
- Safety of physical action
- Mistakes in the physical world are slow, costly and sometimes irreversible. The bar is far higher than for software.
- Measuring progress honestly
- There is no agreed benchmark for taking an idea to a working object. We will need to define one we cannot game.
Current work
Where things stand.
Now
Early- Architecture of the connected engineering representation
- Draft formats for experiment contracts, risk envelopes and evidence packages
- Prototypes of early engineering systems
- Research into a mechanical component pipeline, the first validation domain
- Interfaces with LucidRail for governed physical action
Building toward
Not yet builtA system that moves increasingly far through idea, design, validation and production, with a human deciding at every consequential step.
No customers, facilities or hardware are claimed here. When that changes, this page will say so with specifics.
Design principles
What the system is held to.
- 1
Reality is the reference.
A design is correct when measurements agree, not when a model says so.
- 2
Show the reasoning.
Every decision carries the requirement it serves and the evidence behind it.
- 3
State uncertainty.
Predictions are bands with a source, never bare numbers.
- 4
Failures are data.
A failed test is a result. It is recorded and it changes the design.
- 5
Consequential actions need permission.
Spending, operating equipment and ordering are approved outside Terranoux.
- 6
No silent changes.
A requirement, limit or design choice never moves without a recorded reason.
Working on any of these problems? Write to us.