The Technology Behind Root
Why TCPFN is the right causal AI for process plants.
Root is the first customer-facing technology AI Process Solutions™ (AIPS) supplies to AC+T’s brands. It runs on a temporal causal prior-data fitted network (TCPFN): causal AI that was pretrained once, reads a new plant without being trained on it, and tells engineers how sure it is about every link it finds. This page explains what TCPFN is, how it differs from other AI, and why it fits the way chemical and materials plants actually run.
AI is a Family of Tools
When most people say “AI,” they mean generative AI, the chat tools that write text, code and images. Generative AI is one member of a larger family, and the right tool depends on the problem you are trying to solve.
Generative AI
Creates new content: text, code and images. Our brands use top-tier LLMs every day for research, writing and analysis.
Predictive AI
Learns from past data to forecast what comes next. Credit scores and weather forecasts are everyday examples.
Causal AI
An emerging field that finds cause and effect: what is driving an outcome, and what changes if you act on it.
These core types are combined into AI built for specific jobs. Physical AI runs robots and autonomous vehicles. Conversational AI powers chatbots. Industrial AI works with plant historians, SCADA, MES and PLCs to improve processes. Industrial AI is where Root works.
Why Generative AI Can’t Do the Job by Itself
Wiring a general-purpose chat model to a plant historian is a common first step. Here is how that compares with Root, which pairs time-series analysis with causal AI.
| Generative AI alone, wired to the plant | Root (time series + causal AI) | |
| What it reads | Words: manuals, reports, code and data you upload | Plant sensor data, every tag at once |
| What it finds | Likely causes, based on textbook knowledge | The drivers in your plant’s own data: how big, how soon and how sure |
| Cause or coincidence | Can mistake one for the other | Trained to separate cause from coincidence, with a confidence score on each link |
| A new plant | A new project each time | Pretrained, and works from day one |
| Security | Top-tier models typically run in the provider’s cloud, with plant data sent on every call. Strip out the tag names and the model loses what it reasons from. | Can run on premises, inside the plant’s own network, or work from anonymized data |
| Cost to run | Charged per use, and grows with use | Fixed, on one small model |
| Guardrails | You build approvals, audit and testing | Built in: approvals, audit and verification |
| Time to value | Six months or more to design, build, test and maintain | Weeks: pre-built, pretrained and simple to adapt |
One Situation, Three Answers
Picture an acid plant where SO₂ at the stack keeps creeping up, and no one knows why.
#1
A generative AI chatbot answers “Why is my SO₂ rising?” with the likely causes from textbook knowledge.
#2
A predictive AI forecast answers “What happens next?” with a warning that the plant will likely exceed its permit limit again this month. That is useful, but it comes without a fix.
#3
Root answers “Why, and what should I change?” In the plant’s historian data, SO₂ rises this way after an 8°C drop in bed 1 inlet temperature that follows gas-strength changes. The fix: raise the inlet setpoint with each change.
Most industrial AI falls into one of three groups. Large language models learn from text. Whole-plant models learn from years of one plant’s history so they can predict and optimize it. First-principles models are built by engineers from the equations for the variables they choose to include. Each has its place. None was designed around the question plants ask most often: what caused this?
Many of these models are also custom builds. Each one takes months of data collection and cleanup, fits one unit at one site, and has to be built again for the next. Foundation models changed that pattern in other fields: one pretrained model works on new problems without being rebuilt. Plants needed the same thing for cause and effect. Earlier causal foundation models worked only on static tables of data, with no way to handle readings that unfold over time [1]. Almost all plant data unfolds over time.
Root runs on a temporal causal prior-data fitted network (TCPFN), an architecture published in June 2026 by our technology partner, ProfitOps [1].
Root then links those relationships into ranked, multi-step chains of events, building on the same team’s earlier published work on root cause analysis [2]. The model is also small. The pretrained network is 259 MB, and in the published work it analyzed a 1,275-variable industrial dataset in about six hours on a single GPU [1]. That size is what lets it run inside a plant’s own network.
Trained once, then used as-is.
The model was pretrained on a very large number of synthetic datasets where cause and effect were known. When it reads a plant’s data, it answers in a single pass, with no training, fitting or tuning on that plant. There is no custom model to build and none to retrain.
Trained to say “no effect.”
Much of its training data contains no real relationship at all. That teaches the model to reject noise and shared drivers instead of reporting every correlation it sees.
It reports how sure it is.
For every candidate link, the model estimates whether the effect is real, whether a hidden common cause may be behind it, whether the data can settle the question at all, and whether the link is direct, runs through other variables, or sits inside a feedback loop.
It returns a short list.
On one published benchmark with 104 true relationships, the model returned 160 candidates. A standard Granger causality method returned 6,164 [1]. Ranked by effect strength and confidence, the shorter list puts the few links worth an engineer’s time at the top.
It keeps plant data private.
Root can run on premises, inside the plant’s own network, or work from anonymized data. TCPFN works from the numbers alone, so tag names and units can be stripped out before analysis and matched back afterward on the plant’s side. Top-tier generative AI typically runs in the provider’s cloud, and it reasons from the names and context in the data, so anonymizing the data takes away what it works from.
Why It Fits Process Plants
Chemical and materials plants change all the time. Catalysts age, exchangers foul, feedstocks shift, campaigns switch and units get debottlenecked. A model trained on a plant’s past starts to go stale when the plant changes, and as AI helps plants improve, they change even faster. TCPFN is never trained on the plant, so a change simply becomes part of the data. Move the time window and run it again.
-
Well-run plants hide their own cause and effect.
Good control removes the variation a model needs to learn from. TCPFN says when the data cannot settle a question, so engineers know when a short, targeted plant test is the better next step.
-
Plant data is sensitive.
A historian export shows what a site makes, where its limits are and when it is most exposed. Root can run on premises, with read-only access and no outbound connection, or work from anonymized data, so the plant decides what leaves the site.
-
Findings need someone to act on them.
A short, ranked list with a confidence score on every link fits inside an engineer’s week.
-
Plants can fail from things physics cannot predict.
A valve closes too quickly, a sensor is removed, or an operator doesn't follow procedure. We aren't limiting our observations to ideal theoretical physics because plants don't always behave according to ideal theory.
Your Expertise Stays Yours
Every plant runs on know-how that lives in people’s heads. Experienced operators know which reading matters, which alarm to ignore and what the plant does after a feed change. When they retire, much of that leaves with them. Deloitte and The Manufacturing Institute estimate that U.S. manufacturing could need as many as 3.8 million new employees between 2024 and 2033, and that as many as 1.9 million of those jobs could go unfilled [3].
There is a second, quieter leak. When experts correct a general-purpose AI tool day after day, each correction carries real judgment. With many hosted tools, that learning stays with the provider, and it leaves when the contract does.
TCPFN is built so the knowledge stays with the plant:
-
The model stays the same. The pretrained model does not learn from plant data, and that data is never used to train it.
-
The plant’s knowledge grows. Causal maps, confirmed causes, key operating parameters, and the model context and feedback engineers add all build up for that plant, and each answer gets sharper.
-
A question worth asking any AI provider: if you leave, what happens to everything the system learned from your people?
TCPFN by the Numbers
Published results from the TCPFN research paper [1]:
19
benchmark datasets across five domains, all run with the same model and no tuning
160
candidate links returned on a benchmark with 104 true relationships, against 6,164 from a standard Granger causality method
259 MB
pretrained model
1,275
variables from an industrial plant analyzed in about six hours on a single GPU
In the field, TCPFN identified that recurring sheet breaks at a U.S. pulp and paper mill were tied to an upstream cause the mill’s experts confirmed, with about three hours of median early warning. The operator now knows what to look for and how long to solve it - no capital or expensive consultants.
How TCPFN Finds Cause and Effect
Pretrained once
The model learns from a very large number of datasets where the true causes are known, including many with no real relationship at all.
Reads the plant in one pass
Plant sensor and historian data goes in. There is no training, fitting or tuning on that plant.
Scores every link
For each candidate cause, the model estimates whether the effect is real, whether a hidden common cause may be behind it, whether the data can settle the question, and what kind of link it is.
Builds the causal map
Root chains the strongest links into ranked, multi-step chains of events [2].
Engineers check the physics
AC+T’s process engineers test each link against the process physics, control scheme, and operations manual before anyone changes the plant.
Why AC+T
-
Physics and data on one team.
Through ActiveSite, our engineers build process models, kinetic models and DFT calculations, so every TCPFN finding is checked against the physics.
-
Process engineers behind the technology.
Our people have run plants. They know which problems are worth solving and how to turn a finding into action.
-
Findings connect to fixes.
Our brands supply technologies, catalysts and materials, process engineering and custom equipment, so a confirmed root cause has a direct path to a solution.
-
Secure choices.
Root is SOC 2 compliant and can run on premises or on anonymized data, and one AI use policy covers every brand.
AI Process Solutions: how AIPS supplies AI technologies to AC+T’s companies.
[1] Talupula, S. & Sharma, S. (2026). Temporal Causal Prior-Data Fitted Networks for Panel Data with Learned Reliability Signals. arXiv:2606.20889
[2] Dawoud, A. & Talupula, S. (2025). ProRCA: A Causal Python Package for Actionable Root Cause Analysis in Real-world Business Scenarios. arXiv:2503.01475.
[3] Deloitte and The Manufacturing Institute (April 3, 2024). Manufacturers Need as Many as 3.8 Million New Employees by 2033.
Ai Process Solutions FAQs
What is TCPFN?
TCPFN stands for temporal causal prior-data fitted network. It is a causal AI model pretrained once on synthetic cause-and-effect data, built for time-series data like a plant historian, and published in June 2026 [1].
Does Root need to be trained on our plant’s data?
No. TCPFN reads a new plant in a single pass, with no plant-specific training, tuning or retraining. When the plant changes, you move the time window and run it again.
How is causal AI different from generative or predictive AI?
Generative AI explains likely causes from general knowledge, and predictive AI warns you something is coming. Causal AI finds what is driving the outcome in your own plant’s data, and how sure it is about each link.
How does Root keep plant data secure?
Root is SOC 2 compliant and can run on premises, inside the plant’s own network, or work from anonymized data. TCPFN works from the numbers alone, so tag names and units can be stripped out before analysis and matched back on the plant’s side.
Why does a small model matter?
Small models are easier to secure and operate at fixed costs. The pretrained TCPFN model is 259 MB and ran a 1,275-variable industrial dataset in about six hours on a single GPU [1]. A model that size can run inside a plant’s own network, at a fixed cost.
Large models like LLMs can become expensive quickly, especially as pricing models change and adoption increases.
Talk to us about the Technology
Investors and technology partners: we’re glad to go deeper on TCPFN, how Root is built and where we’re taking it.
AI technology providers: please let us know if you have a technology that can support one of our brands. We are always looking for new tools and partners.
Plant operators interested in Root: reach out to the AC+T brand that serves your industry.