Physics-informed
neural networks

A physics-informed neural network embeds governing equations into the loss function. Train it on data from a physical system and penalise it whenever its output violates conservation of mass, momentum or energy, and it learns a solution that mostly respects the physics rather than one that merely fits the data. For simulation work this is a genuinely powerful idea, and much of the recent progress in scientific machine learning rests on it.

But almost all of that work shares an assumption. The model is trying to reproduce the behaviour of a system it doesn't touch. It predicts, and something else decides.

Why the grid breaks the assumptions

The electrical grid has three properties that sit awkwardly with that setup.

• There's no complete dataset. You can't observe every state variable everywhere. Most of the low-voltage network is instrumented poorly or not at all, so there's no ground truth to fit against in the places where the interesting behaviour happens.
• Events are faster than communication. A frequency disturbance develops in milliseconds. By the time telemetry has reached a central model, the moment to act has gone.
• Actions change the system. The model doesn't only predict, it controls. Every decision reshapes the data distribution it sees next, which breaks the assumption that training data and deployment conditions come from the same place.

That third property is the important one. A model that acts is not a harder version of a model that predicts. It's a different object.

A model of consequences, not solutions

Large Energy Model learns from what happens when something is done. Its training data isn't a set of grid states. It's a set of tuples: the state of the network, the control action taken, the effect that followed, and the context around it.

That changes what the model is for. A PINN approximates the solution to a known equation. This learns the causal relationship between an action and its physical consequence, across many points on a network, over time.

It also solves the data problem, because nobody has an archive of correct grid interventions to learn from. Nobody has been taking them. So the system builds its own record by acting in a controlled way and measuring what follows. No one has to label the outcome. The grid already did.

Where the physics enters

Conservation of energy sits in the training objective, alongside the terms that fit the data. The model is supervised, during learning, against aggregate power and energy balance across the nodes it sees.

This is deliberately a training-time constraint rather than a check applied to outputs afterwards. The distinction is structural. A model that is trained without physical supervision and then has its outputs screened has learned an internal picture of the grid that may be physically wrong, and something downstream is correcting it action by action. A model supervised during training builds a representation that is physically consistent to begin with, which is what makes its predictions about unseen conditions worth anything.

The constraint is a supervision signal, not a proof. It shapes what the model learns rather than guaranteeing any individual output. What it buys is a model whose causal structure reflects the physics of the system it's operating in, rather than one that has memorised correlations and been prevented from acting on the worst of them.

What the model is allowed to do

Physical consistency during learning isn't the same as safety in deployment, and it would be dishonest to present it as such. Those are handled separately.

Authority to adapt is governed by measured stability. The system may continue changing its own behaviour only while the network it's operating in stays within defined bounds. As those margins narrow, its freedom to adapt narrows with them, and if the network approaches its limits, adaptation stops. Permission is earned from measured conditions rather than granted by default.

Implications

Because the model learns at the point of connection, it can run inside the devices that create and consume power. Each node holds a local model. Each is supervised by the same physical relationships. Training can be aggregated across nodes without moving raw measurements off them.

Coordination then happens through the network itself. When one node acts, the others feel it, because they're connected to the same physical system. That's a property of electricity rather than an engineering achievement, but it means a distributed system can stay collectively stable without every node talking to every other one.