Minimisation (or, equivalently, maximisation) of non-linear functions is a widespread tool in astronomy, e.g., maximum likelihood or maximum a-posteriori estimates of model parameters. Training of machine learning models can also be expressed as a minimisation problem (although with some idiosyncrasies). This similarity opens the possibility of re-purposing machine learning software for general minimisation problems in science. I show that PyTorch, a software framework intended primarily for training of neural networks, can easily be applied to general function minimisation in science. I demonstrate this with an example inverse problem, the Out-of-Focus Holography technique for measuring telescope surfaces, where a improvement in time-to-solution of around 300 times is achieved with respect to a conventional NumPy implementation. The software engineering effort needed to achieve this speed is modest, and readability and maintainability are largely unaffected.
Link to PDF (may not be available yet): O1-2.pdf