I describe one of our current major research activities which aims to produce actionable outcomes, namely the prediction of binding free energies of small molecules to proteins. We seek, rapid, accurate and reproducible predictions equipped with uncertainty quantification. In general, it is hard to reliably predict by simulation the outcome of a given scientific process. Faced with such difficulties, scientists today often seek to evade the problem by appealing to machine learning. I look at the advantages and disadvantages of invoking such data-driven approaches. Finally, I discuss our recent discovery that much of the true structure of chaotic dynamical systems is lost on digital computers due to their use of IEEE floating point arithmetic. I illustrate this finding with reference to the generalised Bernoulli map, perhaps the simplest of chaotic dynamical systems. I discuss the consequences of this discovery, inter alia for the application of machine learning to simulation data in “AI systems”.
Building 9 - Lecture hall 2
13:40 - 14:30