Welcome to Apsidion. Why physical AI needs a world model grounded in physics.
What physical AI is missing, why physics is the way to supply it, and how we are building the model.
Language models had their moment when one model learned enough about text to be useful to everyone. Physical AI has not had that moment. We started Apsidion to build the model that brings it: a world model grounded in physics.
Four families of physics. Today each has its own tool; a world model holds them around one core.
What physical AI is missing
Robots, autonomous machines and self-optimizing factories act on the physical world, but they know very little about it. They do not know how a sheet deforms under a press, where the heat goes in a weld, or what happens when a gripper closes on a soft part.
A world model holds that knowledge. Given the state of a process and an action, it predicts what happens next, and how sure it is. We are grounding ours in physics, because physics is the part of the world that is the same for every customer.
Why physics
Engineering computation today forces a choice between trust and speed:
| Speed | Trust | Outside its training data | |
|---|---|---|---|
| Classical solvers | Hours per run | High: built on the conservation laws | Holds |
| Learned surrogates | Milliseconds | No physical guarantees | Breaks |
| A physics world model | Real time | Conservation by construction | Holds |
Conservation of mass, momentum and energy, contact and phase change are the same in a foundry, a robot cell and a human eye. A model that learns them once carries them into every domain. That is what lets it run fast without drifting away from reality, and reason about designs nobody has built yet.
Five gaps
Five gaps separate today's tools from a world model physical AI can use. Most tools close one or two of them. The second and third are not fully closed by any tool today.
| Gap | Where it comes from | What it blocks |
|---|---|---|
| Speed at scale | Serial, stability-limited time steps; attention that explodes across a million elements | Real-time control |
| One representation | Grids and particles cannot hold discontinuities and smooth fields at once | Coupled phenomena |
| Gradients through discontinuities | Tearing, phase change and step changes break differentiation | Inverse design, closed-loop tuning |
| Live calibration | Parameters are fixed offline, with no sensor stream | Adapting to drift, wear and ageing |
| Generalization | Learned operators map functions, not physics | Designs that have never been made |
How the model is built
Because the physics is shared, the core is built once. A new domain brings only its materials, the events that switch its behaviour, its sensors and the actions it can take (Fig. 2).
One shared core. Each domain plugs in its materials, events, sensors and actions.
The core has seven layers, each standing on the one before:
| Layer | What it does |
|---|---|
| 1 · Physical state | Moves between meshes, particles, grids and latent tokens, conserving what must be conserved |
| 2 · Differentiable kernel | Keeps gradients through contact and phase change |
| 3 · Large physics model | Carries mechanisms from one regime to the next |
| 4 · Learned closures | Fill in what the equations leave open, and report their own uncertainty |
| 5 · State estimation | Infers the hidden state of a process from sparse sensors |
| 6 · Action model | Predicts the consequence of an action before it is taken |
| 7 · Inverse reasoning | Works back from a target to a process, and validates it |
The physics is shared across every deployment. Each customer's recipes stay private.
Where it starts
We start in manufacturing. The physics there is dense, the sensors are already on the line, and a better model shows up directly as fewer trial runs.
| How a process gets made | What comes out | |
|---|---|---|
| Today | Experts, separate CAE tools and trial runs, with scrap as the feedback | A process that works after many runs; the knowledge retires with the engineer |
| With a physics world model | The model, plant data and inverse search, calibrated on each line | A validated process in fewer runs, with error bars; the knowledge compounds |
What this blog is for
We will write here about the work as it happens: how one representation holds a crack and a smooth field at once, what calibrating on a live line looks like, and where the model is wrong and by how much. Failures get written up alongside results.