How it works

Don't just predict failures, prevent them by understanding why they happen

Kausalyze builds a causal model of one process unit from 12 to 24 months of historical data, constrains it with process physics and validates the important relationships with your engineers. Use it to trace failures to upstream drivers, test interventions and prescribe changes. First validated root causes in 12 to 16 weeks.

The platform

From plant data to a causal model engineers can use

Historian time-series data, P&IDs or PFDs, lab or quality data and engineering knowledge provide the inputs. Kausalyze proposes time-dependent causal relationships, applies physical constraints and presents the important relationships for engineering review.

Inputs

Historian data

12 to 24 months of time series data.

Process drawings

P&IDs or PFDs to represent the topology.

Engineering knowledge

Process knowledge and a few hours a week during review to inspect, confirm or reject important proposed relationships.

Lab data

Quality samples and assay results, so off-spec product is tied back to the conditions that produced it.

Modelling

Time-aware

Causes may take minutes, hours or days to appear downstream. The model accounts for these delays when linking a driver to its effect.

Causal, not correlational

The model distinguishes variables that drive an outcome from variables that simply move alongside it, using time-series evidence, process constraints and engineering knowledge to pin point true drivers.

Physics-constrained

Mass and energy balances and process connectivity constrain the model to prevent physically impossible solutions.

Engineer-validated

Engineers can review, change, or challenge the model directly, and update it to reflect the plant operation, enabling preservation of institutional knowledge.

Explainable by design

Every conclusion can show the causal path behind it, from upstream driver to downstream outcome, with evidence engineers can inspect.

Applications

Fewer unplanned stops

Identify upstream drivers of recurring trips and assess interventions that could reduce unplanned stops.

Higher yield and throughput

Compare operating windows and set-point changes for their expected effect on quality, production rate and downstream limits.

Lower energy and maintenance

Find the conditions driving excess steam, power use or repeat maintenance, and estimate the value of changing them.

Uses

Find the cause. Test the change. Quantify the impact.

One causal model supports five operational workflows. Inspect the evidence, review important relationships with engineers and compare the expected impact before acting.

Diagnose

Find the cause

Trace a recurring trip, quality problem or process instability back to the variables driving it. Review the ranked upstream drivers, causal path and evidence for each relationship.

Forecast

Forecast what happens next

Estimate when a condition is likely to cross a threshold. Review the operating state, forecast confidence and evidence behind the estimate to plan an intervention.

What if

Test the change before you touch the plant

Change a set-point in the model and inspect the predicted downstream effect. Compare it with leaving conditions unchanged, including the expected effect on process limits.

Act

Prescribe the action

Compare the set-point change, maintenance intervention or operating window most likely to improve the outcome. Each recommendation shows its expected impact and the downstream variables it may affect.

Value

Quantify the value

Use the causal path to connect a proposed change to a measurable plant outcome. Translate expected impact into downtime avoided, production recovered or energy saved. Decisions that translate into real-world savings.

The same model handles
Unplanned shutdowns Recurring quality deviations Throughput loss Rising energy use Abnormal equipment behaviour Chronic process instability
The Industrial OS

From causal model to intelligence layer

Control systems operate the plant. Monitoring systems collect data and flag conditions. Kausalyze provides the causal reasoning layer between those observations and operational decisions.

Find the cause Prevent the failure Optimise the process
Intelligence layerKausalyze · Explain · Predict · Prescribe
Monitoring layerHistorian · Dashboards · Predictive alerts
Control layerSensors · DCS and SCADA · Real-time control

From plant data to operational decisions

Why it's different

This isn't just another analytics tool

System level, not asset-level

Failures rarely respect equipment boundaries. Kausalyze models relationships across connected equipment, process units and shared utilities. The pilot begins with one scoped unit and its relevant upstream drivers.

Explainable and engineer-validated

Every finding comes with the causal path and supporting evidence. Engineers can inspect, confirm or reject the important relationships before acting on them.

Built for process manufacturing

Process structure, engineering knowledge and physical constraints are part of the model rather than an afterthought. Mass and energy balances and process connectivity constrain the proposed causal relationships. Built by process engineers for process engineers.

Sectors

Where we work

Continuous and batch process industries, wherever connected equipment and shared utilities make cause hard to separate from effect.

ChemicalsReactors, separators and heat exchangers
PetrochemicalsCrackers, polymer and derivative units
RefiningDistillation, hydrotreating and FCC
EnergyGas processing, power and steam
PharmaceuticalsBatch and continuous API production
UtilitiesWater, steam and cooling systems
Food and beverageContinuous and batch process lines
The pilot

Four phases, 12 to 16 weeks

The phases
01

Scope and data handoff

Choose one unit and one recurring, costly problem. Agree the target outcome and arrange a secure handoff of historical historian data, tag definitions, engineering units, process drawings and relevant lab or quality data.

You get

  • Agreed unit, problem and success measures
  • Data quality assessment and identified gaps
02

We build the causal model

Use 12 to 24 months of historical data, process drawings and relevant process information to propose time-dependent causal relationships. Apply mass and energy balances and process connectivity as constraints, accounting for the operating conditions represented in the data.

You get

  • A proposed causal map of the unit
  • Supporting evidence, time lags and confidence for review
03

Validate root causes

We review important causal relationships and root-cause findings with your engineers. Inspect the variables, direction, time lag, strength and supporting evidence. Engineers can confirm or reject proposed relationships; the model is updated to reflect the review.

You get

  • Causal map of the unit with evidence per pathway
  • Ranked, engineer-validated root causes for the target problem
  • Remaining uncertainties and data limitations
04

Prescribe changes and quantify value

Identify set-point changes, maintenance interventions or operating windows and test their expected downstream impact in the model. Quantify the value using agreed plant measures, production and cost data. The readout supports a decision about whether and where to expand next.

You get

  • Prescribed changes with expected impact
  • Forecast or time-to-limit estimates where supported by the data
  • Expected commercial value and its assumptions
  • The business case for expansion to further units and plants
Questions

Common questions

How is this different from predictive maintenance?

Predictive maintenance usually focuses on asset condition and the likelihood of failure. Kausalyze models causal relationships across connected equipment and utilities to identify upstream drivers and assess what to change. The pilot starts with one unit and one recurring problem.

Why not use ChatGPT or another language model?

General-purpose language models are useful interfaces for industrial information, but they do not provide the process-specific causal model needed to reason about physical cause and effect. Probablistic predictions are prone to hallucinations, making them impossible to use in decision-critical applications. Kausalyze builds that model from plant data, constrains it with process physics and validates important relationships with engineers.

What data format do you need?

For a pilot, time-series export from your historian for the unit, with the tag list and engineering units. CSV is fine. We confirm the format and sample interval on the call.

Do you need to connect to our historian or control system?

No. The pilot can be built offline from historical historian data, process drawings and relevant lab data. A live-system connection is not required to build and validate the first model.

Who does the work?

Kausalyze engineers build and validate the model. Your process engineer reviews the findings in scheduled sessions and answers questions about the unit. Expect a few hours a week during the review phase.

What if our data has gaps?

We assess gaps, tag changes and periods of bad data at the start. Missing or unreliable measurements may limit which relationships and findings can be supported. The data quality review identifies those limits and any additional information needed.

What do we get at the end of a pilot?

A causal map with supporting evidence, ranked root causes validated with your engineers, prescribed changes with expected impact, and a value assessment based on agreed site measures. The readout includes remaining uncertainties and a recommendation on whether and where to expand.

How is a pilot priced?

Each pilot is scoped and quoted per unit. Ask on the call.

Which unit should we start with?

The one with a recurring problem nobody has fully explained, a well-populated historian, and a known cost per trip. We help you pick it on the call.

What happens after the pilot?

Use the findings and value assessment to decide whether to expand to adjacent units or additional problems. Ongoing data updates and any live-data integration are scoped separately from the pilot.

See what's really driving your downtime.

Bring one recurring plant problem. We will review the unit, available data and the outcome you want to improve, then define a practical pilot scope.

Book a technical review