Skip to content

// Pumps AI Lab

Where pump physics meets machine learning.

First-principles simulation built our training ground. Physics-informed learning is what makes it fast enough to run live on your pump. This is the research engine behind Accel and Sentinel.

Accel + Sentinel

First-principles

Field-hardened

// Simulation-forged, field-hardened

Built on the physics, hardened on reality.

Our models were forged on a corpus of physics simulations spanning every fault mode and fluid regime — now being enriched and validated with real-world operational datasets.

~0
Physics simulations of PD pump behaviour — every fault mode, every fluid regime.
0
Distinct physics modules in the Sentinel engine: fluid state, power, slip, thermal.
0%
Outputs carry a confidence score — certainty shipped as a number, not a claim.

Discipline rule: the number is “about 72,000 simulations” — never inflated, never implying field deployments we don’t have.

// Two methods, one physics

Accel for breadth. Sentinel for depth.

The same governing equations, applied two ways — matched to what each asset’s risk and budget justify.

Accel
Fleet coverage
Neural network

A neural network whose loss function penalises violations of physical equations, so it learns physically consistent solutions. Physics modules stand in for sensors we don’t have — inferring internal state from minimal external signal.

  • Physics compensates for absent sensors
  • Real-time, low-cost, fleet-scalable
  • Confidence-scored inference
Sentinel
Critical depth
Physics engine

Models whose architecture embeds physical laws rather than learning purely from data. Waveform-grade inputs feed a physics engine of distinct modules that explains the subtlest degradations with full root-cause attribution.

  • Physics-derived features, every drift attributable
  • Root-cause mapping in the pump’s actual physics
  • Confidence scores plus RCA

// The quiet hero

The Baseline Performance Curve.

The physics-derived signature of how a specific pump should behave across its operating envelope — established at commissioning, or at manufacture for Twin-at-Birth OEMs. It is the lifetime reference for all intelligence.

Every differentiated claim traces back to it: energy anomalies, fault signatures, slip degradation. Day one, your pump gets its baseline. Every day after, we measure reality against it — drift has nowhere to hide.

// Disambiguation

What we are — and what we’re not.

Our team will be asked “is this like Physical AI?” Getting these distinctions crisp is a credibility moment. We embed exact physics about one thing: PD pumps.

Term
What it actually means
SmartBeam
Physical AI
NVIDIA’s term — AI that perceives, reasons, and acts in the physical world. Robots, embodied agents.
Not us. We don’t actuate anything; we’re an intelligence layer over existing machinery. No robots, no actuation.
World Models / LWMs
Models that learn approximate, general physics of environments from observation.
We embed exact, domain-specific physics from first principles. For a refinery, approximate intuition isn’t enough.
Embodied AI
Agents that learn through a physical or simulated body interacting with environments.
Robotics research. The only body involved is the pump’s — and we read it rather than control it.