Optimising a nuclear fusion reactor with the UK Atomic Energy Authority

Digital Catapult’s first Quantum Technology Access Programme (QTAP) raised awareness, educated end users, and fostered industry partnerships to drive the future adoption and commercialisation of quantum computing. During this first-of-a-kind programme, quantum experts from Digital Catapult and the programme partners ORCA Computing and Riverlane supported participants to explore novel quantum computing use cases.

Can quantum computing help design nuclear fusion reactors by pruning large neural networks?

The UK Atomic Energy Authority (UKAEA) researches fusion energy and related technologies, and aims to position the UK as a leader in sustainable nuclear energy. 

The UKAEA focused on a particular type of artificial neural network known as a Physics Inspired Neural Network (PINN). A PINN uses a classical Machine Learning Neural Network (a programme or model that makes decisions in a similar manner to the human brain) to find solutions to a physics problem formulated as a differential equation in order to leverage the automatic differentiation of modern machine learning packages. Although classical PINNs like this can find valid solutions to physics problems, they can take a very long time to run. UKAEA joined the Quantum Technology Access Programme to understand how a quantum computer could prune the PINN by removing connections between neurons from the network, reducing the size of the model so that it runs more quickly and requires less computing resources.  

What was done?

Digital Catapult quantum computing experts worked with the UKAEA to create a simple PINN to simulate a cooling coffee cup in PyTorch, a machine learning package. An experiment was carried out by controlling which connections were pruned using a binary string pruning mask generated by the ORCA PT-1 photonic quantum boson sampler (quantum computer). Classical optimisation was used to tune the ORCA PT-1 photonic beam splitter angles to change the output until the predictions of the original neural network and the pruned neural network closely matched each other. After pruning half the connections, the PINN still gave reasonable predictions.  

In a separate experiment by UKAEA, a classification neural network that predicted cases of diabetes from different criteria was pruned. After pruning over half the connections the accuracy of the pruned model was almost unchanged. Both these experiments show that quantum computing can successfully reduce the number of connections within a neural network, potentially speeding up classical computing.  

What was learnt?

The simulations show that pruning a neural network using a quantum boson sampler, such as the ORCA PT-1 quantum computer, is a promising approach which UKAEA will investigate further as it could help with the difficult calculations needed to model a nuclear fusion reactor.  

The UKAEA team benefited from learning more about quantum computing, working hands-on with the ORCA SDK software, and the opportunity to work directly with quantum experts to develop their idea of using a quantum computer to prune neural networks.

Be the first to receive news and updates about our work with quantum. Fill out your details in the form below.