Unsupervised ML for Quantum Simulations
Project Description

I was super excited and honoured to have the opportunity to be invited by Rigetti Computing to work on a project at their world-class facility in San Francisco. A few other TKS students and I were able to tour the lab and work alongside some of Rigetti's top quantum engineers. It was amazing to have the chance to get completely immersed into the environment with some of the smartest theorists and engineers in the world, not to mention 6 of the most sophisticated quantum computers ever created.

01. Thought Process

Prior to the event, I had been doing a lot research into drug discovery and the new paradigm created by quantum computers - in the near future, simulating complex molecular interactions won't be a fantasy, and drug discovery will cease to cost hundreds of billions of dollars and years of research. The entire process will likely take no longer than a few days (or even mere hours). Feel free to check out my presentation at TKS to learn more about the future of medicine with quantum computers. With this as our motivation, we developed an ambitious two-part plan for the project; firstly, simulate the Hamiltonian of a helium hydride molecule, and secondly, develop an unsupervised machine learning model to improve upon the standard nelder-mead minimization algorithm.

02. The Project

To understand how to simulate drugs and their molecular interactions with a human system, it's important to first understand their energies. If we can accuractely simulate and model the molecular energy of a molecule those that neighbour it, we'll eventually build up to the simulation of proteins, cells, and maybe one day, an entire human system. Currently limited by the hardware that's available, we're only able to simulate the energy of simple, individual molecules, the simplest of which is helium hydride (HeH+). Our first step was building the simulation - we used a technique known as a VQE (Variational Quantum Eigensolver) to approximate the expectation values of a given Hamiltonian as a function of bond length and angle (graph below). The second part of the project aimed to improve upon the existing industry-standard minimization method for the VQE (called nelder-mead), which becomes unstable and inaccurate with bigger molecules.

03. The Learning Points

Unfortunately, we were unable to finish the second part of the project. We tried using Tensorflow, SciKit Learn, and even a custom minimizing algorithm, but we found that it was a complicated task that would take us a lot longer than we initially thought. Going forward, I'm continuing to work on the project and hope to bring it to completion soon, so stay tuned! The experience was incredibly valuable for a couple of reasons. Working alongside insanely smart people across several industries and disciplines, I learned that the path to working at exceptional companies like Rigetti is not set in stone - it can have a lot of flexiblity, and may not appear clear-cut, but as long as you work hard on things that you love, you'll get there.

Interested? Check out my presentation here! View Presentation