So for my master’s thesis, I decided to do something totally reasonable: teach neural networks quantum physics.
The project is called qubit_dynamics (yes, there’s a typo in the repo name and no, I’m not fixing it — it has character now).
The Problem
Simulating how qubits behave under different control pulses is slow. Like, painfully slow. Traditional quantum simulators (shoutout to QuTiP) are accurate but take their sweet time. When you’re trying to figure out how to control a qubit, waiting around isn’t fun.
The Solution
I built two neural network approaches:
-
PINNs (Physics-Informed Neural Networks) — These networks have the Schrödinger equation baked right into their loss function. They don’t just learn patterns; they learn physics. It’s like teaching a student the actual rules instead of just having them memorize answers.
-
QubiNN — A simpler, data-hungry feedforward network that just learns from examples. No physics, just vibes and gradient descent.
The Results
QubiNN turned out to be ~5000x faster than running QuTiP directly. Five. Thousand. Times.
PINNs were slightly more accurate for predicting control fields, but QubiNN’s speed made it the practical winner for most use cases.
The Stack
PyTorch for the neural networks, QuTiP for generating training data, and a concerning amount of coffee.
Check out the repo if you want to see quantum mechanics and deep learning have a weird but productive friendship.