AI
AI for Industry
Talk about one asset Talk
PHYSICS-BASED PLANT & MACHINE MODELS

We model plants from physics. Calibrated on your data.

Unlock hidden plant capacity, eliminate shift drift, and soft-sense unmeasured process variables in weeks—not years. Our physical models are generated directly from your P&IDs, schematics, and manuals, then fitted to a few weeks of operating data. Every parameter maps to a real piece of equipment, giving process engineers full auditability and operators complete confidence.

Talk about one asset Why physics first
Weeks of Data Fitted to weeks of operating data, not years of historian archives covering every regime.
Open Formats Exported as open FMI / FMU models you can run in Python, Simulink, or your own environment.
100% On-Premises Fully functional on-premises with no outbound connectivity required in the control path.
Client Impact

Plant challenges & the four client impact pillars

Industrial process plants face constant operational friction: shift-to-shift performance variation, conservative setpoints that hide capacity, unmeasured variables that force reliance on slow laboratory samples, and structural questions that historical data cannot answer.

Our calibrated physical modeling approach addresses these challenges directly through four core operational pillars:

Operator–Process Interaction — Eliminating Shift Drift & Human Lag

Industrial process performance varies substantially from shift to shift. Senior operators rely on unwritten intuition developed over decades, while junior operators follow static procedures that struggle during process disturbances. When upsets occur, control loops are frequently switched to manual mode, introducing human lag, over-correction, and hunting cycles with dead-times of 30 to 90 minutes between inlet moves and outlet stabilization.

Our model establishes a continuous, 24/7 digital baseline, creating a clear path from advisory support to full closed-loop automation:

Current Baseline (Operator Dependent) Performance depends entirely on individual operator experience and manual intervention.
Copilot Stage (Advisory Guidance) The model runs continuously, suggesting optimized process settings directly on the operator console.
Autopilot Stage (Autonomous Closed-Loop Control) Once proven, the system writes optimized setpoints directly to the DCS, while operators act as supervisors.
01 / 04
Proven Results

Industry reference cases

Our calibrated physical modeling approach is proven across demanding industrial manufacturing applications.

Case Study

Potato Dry Matter Soft Sensing

Challenge: Variable raw potato dry matter caused starch dosing instability and required slow lab feedback.
Solution: Deployed continuous physical soft sensor for potato dry matter.
Result: Operators receive continuous dry matter data, enabling immediate control of starch addition and eliminating off-spec batches.
Process Control

Toasting Process Control & Quality Optimization

Predicted quality parameters and delivered advisory control settings to operators, achieving smaller quality variation and higher line throughput.

Moisture Precision

Freeze Drying Moisture Precision

Real-time moisture prediction and dynamic control allowed freeze dryers to run closer to accepted moisture tolerances, maximizing yield and cutting drying time.

Yield Recovery

Roasting Process Yield Recovery

Optimized thermal roasting profiles, achieving significant reductions in dry matter loss while keeping all quality parameters within specification.

Energy Balancing

Steam & Energy Demand Balancing

Predicted future steam demand surges across multi-unit operations, enabling optimal real-time balancing between gas boiler and electricity generation.

Core Differentiator

Why physics-based grey-box AI wins

Now that you see the operational impact, the natural question is: Why can’t standard machine learning software or an internal data science team achieve this?

Traditional pure-data machine learning solutions (black-box models) try to infer physical structure strictly from historical data tags. Consequently, black-box models require years of historian data, fail completely when asked to extrapolate outside historic bounds, and deliver opaque statistical correlations that process engineers cannot verify or trust.

Our grey-box modeling approach bridges physics and data. We define the physical model structure upfront using fundamental balance equations, then calibrate parameters on plant data.

“Method is how we build. Consequences are what you buy. Structure carries most of the information, so we identify physical parameters rather than learning an abstract black-box function.”
Dimension Black-Box AI (Competitors) Grey-Box Model (Our Solution) Direct Difference for Your Plant
Data History Needed Requires years of historian data covering every operational regime. Requires a few weeks of representative operating data. Launch projects immediately without waiting for years of clean data archives.
Extrapolation Power Cannot predict performance outside observed historical data. Extrapolates reliably within physical validity limits. Test higher throughput, new recipes, alternative feedstocks, or fouled states safely.
Explainability & Trust Black-box statistical correlations; no physical meaning. Parameters map to physical quantities on named equipment. Engineers can audit parameters; operators trust and act on recommendations.
Fleet Transferability Must re-learn every new machine from scratch (100% rework). Calibrated model of machine #1 forms ~80% of machine #2. Rolling out across 20 fleet units becomes dramatically faster and cheaper.
Methodology

Automated 4-step methodology: how a model gets built

We automated the highest-cost phase of physical modeling: converting engineering documentation into computational model graphs.

1

Your documentation becomes a graph

P&IDs, schematics, datasheets, operating manuals, and control narratives are ingested into an automated system that extracts a machine-readable graph of equipment, connections, and instruments. The graph is yours to keep for maintenance, change management, and drawing verification.

2

The graph becomes physical equations

Components are instantiated from open modeling libraries, closing mass, momentum, and energy balances. Formulations are structured specifically to behave well under parameter estimation.

3

The equations are calibrated on your plant

Parameters are fitted against operating data with balance reconciliation. This step routinely exposes instruments that have been quietly miscalibrated for months or years. If a parameter cannot be uniquely identified from available data, we state the error band and define the shortest test to resolve it.

4

One day on-site, remotely maintained

We conduct a single one-day site walk to capture operator knowledge—unrecorded valve quirks, untrusted readings, grade-change habits. The remaining work is executed remotely over a secure data connection.

Scope & Safety

Scope, safety, and open delivery formats

Regulatory control stays in your DCS. Models are delivered in open formats you can run without vendor lock-in, and maintained as the plant changes.

Regulatory control remains in DCS

We write setpoints to your DCS. Every interlock, safety limit, and permissive stays in the DCS, which decides whether to accept setpoint updates.

Zero Safety Instrumented System (SIS) touch

Safety systems are separate, out of scope, and untouched.

On-premises cybersecurity

Fully functional on-premises with no outbound cloud connectivity required in the control path. Built following IEC 62443 zoning standards.

No vendor lock-in

Models are delivered as standard Functional Mock-up Units (FMUs) compliant with DEXPI, OPC UA, ISA-95, and FMI standards. You can run them in Python, Simulink, or your own environment without ongoing proprietary software reliance. Source and model escrow are standard.

Subscription-based revalidation

Plants change, sensors drift, and catalysts age. We continuously monitor drift, revalidate models on a schedule, and report instrument health as a direct by-product of running the model.

Discovery Track & Contact

Working with us: the Discovery Track

We start with a low-barrier Discovery Track focused on a single asset before asking for broader project commitments.

01

Discovery Track

Client input required

  • Plant P&IDs / schematics
  • Few weeks of read-only operating data
  • Process engineer point of contact
  • One day on-site access

Client deliverable & outcome

  • Calibrated physical asset model
  • Validated soft sensor / optimization report
  • Quantified ROI and acceptance metrics
02

Advisory Deployment

Client input required

  • Approved DCS setpoint interface
  • Operator feedback during shift runs

Client deliverable & outcome

  • Real-time advisory copilot on operator console
  • Side-by-side comparison with historical operator choices
03

Closed-Loop Authority

Client input required

  • Formal client sign-off on error bands

Client deliverable & outcome

  • Automated DCS setpoint updates
  • Continuous model maintenance & revalidation subscription
Target Sectors

Where we work

  • Food and beverage processing
  • Specialty and fine chemicals
  • Industrial utilities and thermal plant
  • Industrial heat pumps and heat recovery
  • Cooling and data-centre thermal management
  • Engineering contractors (at capital project design stage)

Ready to optimize your first asset?

The most productive first conversation focuses on a single machine or process unit with one clear problem.

Email

PLACEHOLDER@example.com

Telephone

+00 000 00 00 00

Location

PLACEHOLDER address
Belgium

Send us your P&ID or schematic and tell us what you wish you could measure. We will tell you whether our physical modeling approach fits before anyone writes a proposal.