01 — SENSE
Every node feels the grid.
Voltage, current, frequency, rate of change of frequency, power factor. Measured continuously at the point of connection and precisely timestamped. Each connection point becomes an instrument, not just a meter.
LARGE
ENERGY
MODEL — Physics-supervised intelligence for the grid
Physics-supervised intelligence for the electrical grid.
THE PROBLEM
The grid now moves faster than the cloud.
A frequency disturbance can develop in milliseconds. Sending every decision to a distant data centre adds latency exactly when the grid can least afford it. Some decisions have to be made where the electricity is.
WHERE IT STARTS
And the disturbance starts at the edge.
EV chargers, heat pumps, solar inverters and induction hobs are changing the electrical behaviour of millions of buildings. Much of that activity sits behind the meter, beyond the grid operator's real-time view. Yet in aggregate, it becomes part of the grid.
THE TURN
So the intelligence has to live where the physics happens.
Edge appliances and devices that are changing the behaviour of the grid can also help stabilise it. If they can sense, decide and act locally, they can respond on the timescale of the event itself.
THE OBJECTION
But infrastructure cannot run on a model that guesses.
A data-driven model can produce an answer that looks plausible but violates the physical limits of the system. In critical infrastructure, checking the answer afterwards is not enough. The physics needs to be part of how the model learns.
LARGE ENERGY MODEL
So physics does the supervising.
Large Energy Model predicts how the grid will respond, takes a controlled action, and measures what actually happened. Conservation of energy is built into the training process, constraining what the model can learn. The grid itself provides the feedback.
Large Energy Model + PAVLINA
Every one of these devices already has a computer inside it.
An EV charger runs firmware. An inverter runs a control loop. A heat pump has a processor deciding when the compressor turns on. The silicon is already installed, in millions of units, sitting exactly where the physics happens. It just isn't being asked to do anything about the grid.
That's what's new. Ten years ago these devices were few, small and dumb. Now they're numerous, powerful enough to run a model, and collectively large enough to move the network. The problem and the means to solve it arrived together.
Large Energy Model is the intelligence layer that learns the physical system.
PAVLINA is the control architecture that turns that understanding into coordinated action across it.
Neural decision map
Forecast demand ramp exceeded local headroom, so storage discharged before the constraint bound.
Topology-aware dispatch routing
Discharge storage ahead of the evening ramp, routed hop by hop to the assets that can act on it.
02 — ACT
And pushes back on it.
Each node has its own energy buffer and can respond directly to changing grid conditions. Decisions are made locally, without waiting for instructions from the cloud or drawing on a connected vehicle's battery.
03 — LEARN
Every action creates feedback.
Predicting consequences, taking controlled actions, measuring what actually happened, updating on the differences. No one has to tell Large Energy Model whether it was right, because the grid already did.
04 — GOVERN
Authority is earned, not assumed.
Adapting would only happen when stability margins remain within defined bounds. As those margins narrow, its freedom to change its behaviour narrows with them. If the network approaches its limits, adaptation stops.
WHAT IT AMOUNTS TO
A grid that learns from itself.
Instead of observing the grid from a distance, Large Energy Model becomes part of it. Distributed nodes sense local conditions, respond within defined limits and learn from what happens next. Together, those local interactions allow the network to adapt without every decision being sent back to a central control room.
Electricity is the first application. The same principle extends to other physical systems: when AI is given authority to act, the laws governing that system can also constrain how it learns.
SIMULATED SCENARIO · NOT LIVE CONTROL
- FREQUENCY
- 49.988 Hz
- DEMAND
- 33.4 GW
- STABILITY
- 97
- RESPONDING
- 0 / 10,000
- NO ACTIVE EVENTS
System inside envelope. Select a disturbance.
POSITION
What already exists.
4 UK patents filed, 1 PCT filed, 2 registered trademarks, and a simulation demo. All built and paid for without outside investment.
PATENTS
4 UK patents pending
GB2519693.2
GB2519709.6
GB2602143.6
GB2608979.7
TRADEMARKS
Registered
Large Energy Model®
PAVLINA®
CAPITAL
Self-funded to date
Core IP developed independently
NEXT BUILD
From simulation to the first physical node.
The next step is to take the architecture beyond simulation and into a controlled electrical testbed. The first node will bring sensing, computation and controllable power hardware together, allowing it to observe local electrical conditions, predict how the system will respond to a bounded intervention, act through an inverter and energy buffer, and compare the outcome with what it expected.
FOUNDER
Engineering the grid more like a living system.
Aaron Tan
FOUNDER
Stabilising planetary-scale intelligence.
MBBS + PhD, University College London. Previously visiting scholar at Stanford University. Google Scholar
- Citations
- 6,800+
- more than half since 2021
- Research impact
- h-index 34