Workshop 

Numerical unmixing of magnetic components – success and pitfalls after 29 years

​

Ramon Egli and Eric Font.

 

The identification of magnetic mineral components in rocks and sediments is one of the most important tasks in paleo-, rock- and environmental mechanism, especially for last developments in these fields, where the interpretation of paleomagnetic signals and magnetic proxies for geologic and environmental processes relies on precise identification of distinct groups of magnetic minerals – the so-called magnetic components – with a common origin. The successful application of numerical unmixing techniques produced a paradigm shift from simple rock magnetic analyses in terms of magnetic mineralogy and bulk grain size to mixtures of discrete magnetic components. Numerical unmixing techniques were for instance pivotal for the identification of magnetofossils – now being recognized as a widespread magnetic component in freshwater and marine sediments with important implications for relative paleointensity records and environmental proxies. Magnetic unmixing methods are based on principal component analysis (PCA), analysis of magnetization curves, including first-order reversal curves (FORC), and a combination of both. The analysis of magnetization curves depends in turn on the concept of coercivity distribution, in which case each magnetic component is represented by one distribution. Coercivity distributions of magnetic components with the same mineralogy are usually highly overlapped and therefore difficult to disentangle. Parametric unmixing approaches model each coercivity distribution with one or more model functions, while non-parametric approaches are based on some mathematic decomposition methods similar to PCA. In both cases results are negatively affected by measurement noise and by slight variations of the magnetic properties of individual components. We show apparently simple cases with two components known a-priori where numerical unmixing methods fail to produce the correct results if not correctly implemented. Starting with these examples we discuss general pitfalls of numerical unmixing methods and ways to obtain more stable solutions and check their reliability, along with various applications of numerical unmixing techniques. We also show how additional information, direct observation of magnetic particles, and forward modelling can help in gaining a better understanding on the physical nature of magnetic components.