STEM, Research

Dark matter simulations using quantum neural networks

Among all the mysteries of our universe, dark matter is a particularly stubborn one: we can’t see it, touch it or interact with it through the electromagnetic field. What we can do is use equations telling us how it might behave. The problem ? They’re very difficult to solve efficiently

Our current understanding of the universe includes the existence of a special type of matter that makes up 27% of the cosmos' mass while remaining completely invisible to us - hence the name: dark matter. As it has never been directly observed, we use theoretical models to infer what its properties might be. One such model, known as Fuzzy Dark Matter (FDM), suggests that this invisible substance behaves like some sort of quantum wave stretched on the scale of galaxies. As part of my research at the LASTRO laboratory, I became interested in this model and decided to try finding solutions to the equation governing it, known as the Schrödinger-Poisson equation, given by:

The main problem is that as soon as you try to describe a somewhat complicated system, it gets almost impossible to solve the associated differential equations with a pen and paper (as any student who's taken a physics class could tell you). That’s why I'm trying to combine two developping technologies to solve them:

  • The first tool is something called Physics-Informed Neural Networks (PINNs). These are machine learning models that work similarly to regular neural networks - systems that learn patterns from data- but with one key difference: they’re trained to obey the laws of physics. Instead of learning blindly from examples, PINNs are guided by the equations that describe how things behave in the real world. This helps ensure that the results make physical sense and might actually reflect how FDM behaves. It also means we don’t need huge amounts of training data, since the model already has an "intuition" for the physical laws

  • The second tool is quantum computing, which takes advantage of the strange rules of quantum physics - like superposition and entanglement - to process information in new ways. Unlike classical computers that work with bits (0s and 1s), quantum computers use quantum bits which, in my case are represented by photons (the particles that make up light). These can perform certain types of calculations far faster than even the best supercomputers, especially when simulating systems governed by quantum mechanics.

So how do we combine them? Instead of using a standard neural network with classical operations, we build a Quantum Physics-Informed Neural Network (QPINN). That means we replace the usual parts of the network with operations performed on a quantum circuit. The input is encoded into quantum states, which then evolve through a sequence of quantum gates (think of them as the quantum equivalent of the logic gates you'd find in a regular computer). The final output is measured and compared against what the physics tells us it should be. We adjust the quantum circuit parameters until the output matches our desired result, effectively training the network to learn the solution to a physical equation. This is just like what happens in PINNs, but powered by quantum hardware.

By combining these two powerful tool (physics-informed learning and quantum computing), we’re building a new way to solve equations that are otherwise at best very slow and difficult to solve. This approach doesn’t just help us simulate fuzzy dark matter, as it opens the door to tackling a wide range of complex problems in physics. If we want to understand the invisible structure of our universe, we need methods that are just as bold and unconventional as the questions we’re trying to answer.