Abstract

Materials science differs from the pure subjects in that it attempts problems at the level of complexity that is posed, rather than by simplification to study a narrow aspect. This raises challenges that usually involve interdisciplinary skills and myriads of non-linearly interacting variables. A second distinction is that there is a genuine yearning to validate predictions. In this lecture I will introduce the method of neural networks within a Bayesian framework, a method that is a form of mathematical modelling that can help resolve complexity whilst striving for broader solutions. I will demonstrate that the method permits the discovery of new phenomena, and the quantitative design of new materials with a minimum use of resources. At the same time, it introduces a culture in which both noise and modelling uncertainties are considered in order to realise the value and limitations of the mathematical approach. Some recent successes in the design of new materials include the -TRIP steel, a welding alloy that cancels the development of residual stress, and a nickel alloy that is cheap enough to serve in ultra-supercritical steam driven power plant.