Pretrained on the physics of your process, it reads your data and tells you what a change will do, and why.
Unlike a digital twinNo model to build or calibrate for your site. Ceptron is pretrained once per domain, and your data places your site among millions of simulated ones.
The causal model ships first: Temporal Causal Foundation Model, open weights on Hugging Face.
A simulator alone needs an engineer to calibrate it. A learned model alone cannot say why. Ceptron combines physics, a world model and a causal model, so it calibrates itself from your data and explains its answers.
ProcessAny operation where material or work flows through connected steps, with delays and control: a mill, a warehouse, a GPU cluster.
The rules of your kind of process: what flows in, what flows out or builds up, how fast each step can run, and how long each delay is. They set the ceiling. No fitted numbers.
Pretrained on millions of simulated worlds, so it predicts a change you have not made yet, with a confidence band.
Reads your data, finds what drives what and at what delay, and names the lever to pull.
Knows what can happen, what will happen if you change something, and why it moved. Calibrated by your data, not by hand.
Three strata, one question. Your AI agent, or you, asks in plain language, with the data you already collect, live or recorded. The Temporal Causal Foundation Model reads your data and finds where you are, the Process World Model predicts the change, and physics bounds the answer. Scroll, or click a step.
Your AI agent, or anyone on your team, asks in plain language: what happens if I change this, what should I set, which upgrade pays back, why did it miss. Connect what you already collect, live or recorded: sensor data, logs and event records. Nothing new is installed, and no experiments are run on the live operation.
Balances for every step, the rate each step can run at, and the delays between steps. Each domain has its own: chemistry and heat in a mill, queues and travel times in a warehouse, compute and memory in a GPU cluster. The rules and the ceiling that every answer must respect, with no fitted numbers: when a prediction misses, the fix is a missing piece of physics.
Pretrained once on millions of simulated histories, one for every kind of site. It predicts a change you have not made yet, in one pass, with a confidence band. How the Process World Model works →
Our temporal causal foundation model reads your data, live or recorded: sensor data, logs and event records. It names what drives what, at what delay and by how much, and tells the world model which of its simulated worlds is yours. Where a tag is missing, it says so. Its discovery head is the open model.
The setting to change, what it will do, with a confidence band, and the cause behind what you have been seeing. Every answer is explained by the Temporal Causal Foundation Model, predicted by the Process World Model and bounded by physics. People approve every change.
Ask in plain language. Every answer comes with the numbers and how sure the model is. People approve every change.
A process world model: physics that says what can happen, a world model pretrained on millions of simulated worlds that says what will happen if you change something, and a causal model that reads your data and says why it moved.
A twin or simulator needs an engineer to calibrate it per site and cannot say why something moved. Ceptron calibrates itself from the data you already collect, answers what-if before anyone touches the line, and names the cause when reality differs.
No. Ceptron is pretrained on simulated worlds. Your data, live or recorded, is read when you ask a question, so there are no experiments on the running operation.
Pulp and paper mills, NeoClouds, and, through the causal model, chemicals, feed mills and warehouse automation. PFN Studio extends it to your own operation.
The causal model inside Ceptron. It reads plant data and reports which variables drive which, at what delay and by how much. It also runs on its own on any historian.
The causal model can be run now: the Temporal Causal Foundation Model, with open weights on Hugging Face and code on GitHub. Install the package, load the model, and call discover on a time-series table. The world model runs with us.
The weights are CC BY-NC 4.0: free for research, teaching and personal projects. Commercial use needs a separate license; book a time.
The studio Ceptron is built with: describe an operation, generate simulated worlds from the description, pretrain, evaluate and deploy. Your team can use it to build a process world model of your own operation.
Public open-weight checkpoints, including mixture-of-experts and hybrid-attention designs, on datacenter GPUs from NVIDIA and AMD, newer accelerators, workstation cards and local machines.
Tell us what you want to model and we will be in touch.