Lab cadence is too slow
Manual sampling and lab turnaround leave long blind spots between readings, exactly when a process is drifting toward the spec limit.
Know where quality is heading, not just where it's been.
Your lab tells you what already happened. ProcessMiner predicts quality every 30 seconds, by grade, so your team can correct course before a deviation becomes scrap.
The problem with lab-based quality
By the time the lab result comes back, the batch is already made.
Most quality checks happen after the fact: a sample pulled off the line, tested, and reported back minutes or hours later. That gap is where scrap, rework, and off-spec product get made. Operators are firefighters; they don't have time to optimize the process every 15 minutes on their own.
Grade changes reset the picture
Every product grade has its own quality limits. A single generic model can't keep up as production shifts from grade to grade throughout the shift.
A single bad reading isn't the goal
Chasing an exact point value distracts from what matters: the direction the process is trending, and whether it's headed out of bounds.
How real-time quality prediction works
A live quality prediction, running continuously between lab tests
The model acts as a soft sensor. It learns your process from historical data, then predicts quality continuously, filling the gap between lab samples with a live, trend-based estimate that updates as conditions change. We focus on trend accuracy over point accuracy because that's what actually drives action: telling you the direction quality is moving well before a lab result would, not guessing a lab value to the decimal.
Proactive optimization
Prevent before it happens
ProcessMiner predicts the trend and corrects in real time.
- Quality prediction refresh. Updates every 30 seconds, continuously.
- Recommendation cadence. Every 15 minutes, only when the trend is out of bounds.
- Grade/SKU-specific limits. Applied automatically as production moves grade/SKU to grade/SKU.
Inside the prediction engine
What keeps quality predictions accurate and current
Real-time prediction isn't one model; it's a system that keeps testing, retraining, and adapting so the prediction in front of your team is always built on the best available fit.
01 / Grade-specific modeling
Built around your product grades, not a generic average
Quality limits, upper, lower, and target, are defined per product grade. As production moves from one grade to the next, the platform applies the correct limits and model automatically, without a manual reset.
- Manages 5 to 40+ grades without manual reconfiguration
- Built-in grade-transition logic adjusts the active model as the product changes
02 / Automatic model selection
Hundreds of approaches tested, the best fit wins
The platform tests hundreds of modeling approaches, from statistical regressions to random forest and deep-learning networks, and automatically switches methods when a model's performance degrades. You don't have to pick a technique or babysit a model.
- Auto-switches among modeling methods when performance drops
- Multiple parallel models per quality metric, combined to avoid trade-offs across metrics
- Continuously retrains on every new lab value, so predictions stay anchored to how the process behaves today (min. 6 months of historical data per metric, 6–12 preferred)
Reactive vs. proactive
03 / Connected to the rest of the platform
Prediction is the input, not the endpoint
A prediction on its own tells you something is drifting. Paired with the rest of the platform, it tells your team what to do about it: setpoint guidance from Intelligent Centerlining, and a ranked list of contributing factors from root cause analysis.
- Feeds Intelligent Centerlining with the live trend needed for setpoint guidance
- Triggers root cause analysis the moment a prediction crosses a grade limit
Prediction feeds the platform
The Process Keeps Getting Better
"Every time a new lab result comes in, the model learns from it. That's what keeps a live prediction honest between tests, not just a guess running on a timer."
— Karim Pourak, CEO
30s
Prediction cadence
Continuous, live
15min
Recommendation cadence
Only when out of bounds
$400–500K
Value of a 0.1% quality gain
Per year, average paper mill
$2.4M
Value per machine, per year
50% attributed to the platform
Figures reflect actual ProcessMiner deployments in pulp and paper. Results vary by facility, process, and data quality; past performance is not a guarantee of future results.
See the prediction engine on your process
See it on your process, or start with a pilot.
We'll walk through how the model would apply to your grades, your historian, and your quality metrics, then decide together whether a pilot makes sense.
Already seen the platform in action? If you're ready to start a pilot, click below.
~30 min
to connect your historian
Hours
to a first live prediction model
Under 2 weeks
to full deployment
Questions, answered
Frequently asked questions
What does "real-time" actually mean here?
Quality predictions update every 30 seconds across every grade-specific parameter you are tracking. Corrective recommendations are issued every 15 minutes, and only when the process is trending out of bounds, not as a continuous stream of alerts.
How accurate are the predictions?
We focus on trend accuracy because that is what actually drives action. The models are built to tell you where quality is heading, not to hit an exact lab value to the decimal.
What types of predictive models does PM Studio use?
The platform tests a vast library of hundreds of models, ranging from basic statistical linear regressions to advanced machine-learning and deep-learning neural networks, and automatically selects the best-performing model for your process.
How does the system choose the best-fit model?
The platform automatically picks the best model based on predefined performance criteria, and auto-switches to a different modeling method if performance degrades. If your domain experts have preferences for certain machines, grades, or mills, the selection criteria are fully adjustable.
Can the system handle frequent grade changes or product transitions?
Yes. The platform has built-in grade-transition logic that continuously adapts the model based on the active grade, and can manage anywhere from 5 to 40+ grades without manual reconfiguration.
How much historical data do we need before this works?
A minimum of 6 months of historical data per quality metric, with 6 to 12 months preferred. That history is what the model uses to learn your process before it starts predicting live.
Get in touch
Have a question?
Whether it's our platform, your process, or a free in-depth process data analysis and pilot you'd like to see in action, let us know.