Forward problem
Causal inference
Given a known causal model, estimate the effect of a change. How much does changing X move Y?
How it's built
A causal model of your system that engineers can read, question, and correct.
Rhizant drafts the graph from your documents, confirms it against your data, and builds in the physics you already know. Every link shows its evidence and its confidence.
Inputs and outputs
Sensor data and manuals in. A twin you can ask why out.
The real system streams sensor and operating-condition data into the causal digital twin. Engineering documents and, if present, an existing digital twin also feed the model. Engineers get answers back, or the twin acts in closed loop.
Inputs
- Real-time sensor dataStreamed from the real system
- Environmental dataOperating conditions
- Engineering documentsManuals, schematics, maintenance logs
- An existing digital twinIf you have one
Causal digital twin
Keeps only the decision-relevant causal links
Outputs to engineers
- Recommended actions
- Root cause analysis
- Intervention analysisPredicted effect of a proposed change
- Outlier analysisFlagging abnormal behavior
- Causal strengthsHow strongly each factor drives another
Digital twin vs causal digital twin
With or without a digital twin
A digital twin is a virtual replica of a physical asset, kept in sync with live sensor data and models. It tells you what is happening. A causal digital twin adds the cause-and-effect structure, so it also tells you why, and it can simulate interventions.
They are complementary
The digital twin tracks thousands of variables at high frequency. The causal twin keeps only the decision-relevant causal links needed to simulate interventions.
No digital twin needed to start
You do not need an existing digital twin to get value. If you have one, the causal twin sits on top of it.
Structural causal models
Two core tasks
Correlation has no direction; causation does. Causality is represented with structural causal models: a directed acyclic graph showing which variables cause which, plus structural equations describing each link. These equations are one-way; they cannot be inverted like ordinary algebra.
Inverse problem
Causal discovery
Learn the causal graph itself from data. This is hard because many different causal models can produce the same data, so it needs assumptions plus extra techniques.
Time matters
A cause always comes before its effect.
That is why time-series sensor data makes causal analysis stronger.
It reveals recurring cause-and-effect patterns and the time lags between actions and outcomes, and it helps detect hidden (unmeasured) factors that influence the system.
Three steps
Draft. Confirm. Add the physics.
01
Draft the graph
The manual already knows most of the causal structure.
A language model reads manuals, schematics, and maintenance logs and proposes the first causal graph.
02
Confirm with data
Your sensors don't lie about order in time.
Time-series causal discovery confirms, corrects, and weights each link, finds time lags, and flags hidden factors. Each edge gets a confidence score.
03
Add the physics
Equations you trust become part of the model.
Known physical relationships are written in as structural equations, so the twin needs less data and stays plausible in conditions it has never seen.
Ask the graph
Every link lists its source and its confidence.
Engineer
Which links in this graph came from the manual, and which from the data?
Rhizant
Fourteen links were drafted from the service manual and the maintenance log. Eleven were confirmed by the sensor data, two were corrected (the direction was reversed), and one was removed because the data showed no effect.
Three new links were found in the data that the manual did not mention. Each link lists its source and its confidence.
- 14 drafted from documents
- 11 confirmed by data
- 2 direction corrected
- 1 removed
- 3 new, found in data
Open problems
The open problems in causal digital twins
These are the problems that keep causal digital twins from working in practice today. The approach is built around them, not around a generic causal-inference library.
-
01
Automation
Building causal digital twins today takes a lot of manual, ad hoc expert work: building models, integrating data sources, and interpreting results.
-
02
Real-time adaptability
Twins need to update their causal models on the fly as new conditions or fault modes appear, despite noisy data and changing system structure.
-
03
Scalability
Causal discovery gets much harder in complex systems with many variables. Scalable, continuous causal learning is still an open problem.
-
04
Validation and trust
A twin may drive critical decisions, like shutting down a machine. But checking a causal model is harder than checking prediction accuracy: counterfactual predictions often can only be confirmed through experiments or long-term observation.
-
05
Combining expert knowledge with data
Finding the right balance between engineers' domain knowledge and causal structure learned from data, for example with scalable Bayesian approaches.
How Rhizant addresses them
LLM-drafted causal graphs
Large language models read manuals, schematics, and maintenance logs and draft a first causal graph, which the data then confirms or corrects. This replaces weeks of manual expert work.
Confidence on every answer
Every diagnosis and what-if answer comes with how sure the model is, using Bayesian uncertainty. Engineers know when to act on it and when to double-check, which matters before decisions like shutting down a machine.
Physics + data hybrid models
Known physics equations are built directly into the causal model as structural equations. The model needs less data, learns faster in high-dimensional systems, and stays physically plausible in situations it has not seen.
Inspectable graphs
The causal graph is shown edge by edge, with the evidence for each link. Engineers can question, override, or add links from their own knowledge.
Input types
Uses the sensor data and documents you already have.
No new experiments are needed to start. Nothing has to be centralized first.
- 01Process historians and SCADA/DCS tags
- 02Test-cell and telematics streams
- 03Maintenance and work-order logs
- 04Manuals, schematics, and P&IDs
- 05Existing digital twins and simulation models
- 06Operating-condition and environmental data
See it drafted from your own manual.
Bring one machine and one unexplained fault. We will walk through what its causal graph would look like.