Machine learning applied to rock geochemistry for predictive outcomes: The Neapolitan volcanic history case

Highlights

• We explore machine learning in analysis of rock geochemistry from the Neapolitan volcanoes.

• We use a data set of 9800 volcanic rocks assembled from the literature.

• Predictive ability has been evaluated for several machine learning methods.

• Machine learning prediction can be useful in petrology and tephrostratigraphy.

• Machine learning is an interesting tool to improve volcanological studies.

Abstract

In this paper we explore the efficiency of various machine learning techniques to determine the volcano source, the eruptive formation and the eruption period of volcanic rocks when their chemical contents are known. With this aim, we assembled a data set of 9800 volcanic rocks from the open-access literature. The rocks belong to eruptive formations from Somma-Vesuvius, Campi Flegrei, Ischia and Procida volcanoes, in the Neapolitan region of Italy. The data set includes content of major oxides and trace elements, as well as Sr and Nd isotope ratios, eruptive periods, eruption formations and volcano source. Some discrete numerical variables are missing in certain samples resulting in data exclusion and measurement inhomogeneity. Our results indicate that, despite such issues, some machine learning algorithms have a very high prediction ability, i.e., at >70%. The achieved results are interesting in order to facilitate the managing of new data for volcanological reconstruction and tephrostratigraphic studies.

Introduction

Volcanic rocks are continuously analyzed in order to understand the origin, evolution, and behavior of magmatic and volcano systems with assessment of environments and natural hazards. The literature allows several data that characterize rocks produced by each eruption and a lot of data often exist for various recognized eruptive deposits. Rock geochemistry studies provide the basilar characterization of eruptive deposits and of related frozen-in magmatic processes. An example that has international audience and we particularly know well, is from the Neapolitan volcanoes (Fig. 1) where rock geochemistry studies are really abundant (Arienzo et al., 2009, Arienzo et al., 2016; Barberi et al., 1978; Beccaluva et al., 1991; Belkin et al., 1993; Brown et al., 2008, Brown et al., 2014; Cannatelli et al., 2007; Caprarelli et al., 1993; Carbone et al., 1984; Casalini et al., 2017; Cecchetti et al., 2003; Civetta et al., 1991; Civetta and Santacroce, 1992; Civetta et al., 1997; Cioni et al., 1995; Cortini and Hermes, 1981; Crisci et al., 1989; D'Antonio, 1991; D'Antonio and Di Girolamo, 1994; De Astis et al., 2004; De Vita et al., 1999; D'Antonio et al., 1999; Di Renzo et al., 2007; Fedele et al., 2008, Fedele et al., 2016; Fourmentraux et al., 2012; Forni et al., 2016, Forni et al., 2018; Gebauer et al., 2014; Hawkesworth and Vollmer, 1979; Insinga et al., 2006; Iovine et al., 2017; Landi et al., 1999; Marianelli et al., 1999, Marianelli et al., 2006; Mastrolorenzo et al., 1993; Melluso et al., 2012, Melluso et al., 2014; Pabst et al., 2008; Orsi et al., 1992, Orsi et al., 1995; Pappalardo, 1994; Pappalardo et al., 1999, Pappalardo et al., 2002a, Pappalardo et al., 2002b, Pappalardo et al., 2008; Piochi, 1994; Piochi et al., 1999, Piochi et al., 2005b, Piochi et al., 2006, Piochi et al., 2008; Rosi et al., 1993; Santacroce, 1987; Santacroce et al., 1993, Santacroce et al., 2008; Signorelli et al., 1999a, Signorelli et al., 1999b; Smith et al., 2011; Somma et al., 2001; Stock et al., 2018; Tomlinson et al., 2012; Tonarini et al., 2009; Vezzoli, 1988; Villemant et al., 1993; Webster et al., 2003).

Traditionally, these studies use bivariate plots, sometimes triplots, to recognize evolution patterns and to interpret data by means of visual inspection (Harker, 1909; Pearce, 1968; Le Bas et al., 1986; Rock and Carroll, 1989). However, data abundance may determine ambiguous data clouds, numerous samples cannot be efficiently compared and several variables are not certainly managed. Therefore, traditional methods are not practical for large databases and are subject to a certain arbitrary selection of discriminant variables by operators. Tough, statistical approaches (discriminant analysis, clustering, etc) are used in rock geochemistry (e.g. Buccianti et al., 2006; Verma et al., 2005) although they are not fully adequate in volcanic rocks showing compositional variability and give more satisfactory results in other branches of the Earth Sciences, for example the fluid geochemistry. In any case, according to our knowledge, statistical approaches have not been used for geochemistry of Neapolitan volcanic rocks, the object of this study. Machine learning (ML) methods provide effective tools for big data analytics. The methods are increasingly applied in Earth sciences, i.e., petroleum, geochemical prospecting, mineral resource, geochemistry (Agoubi et al., 2016; Bolandi et al., 2015; Zuo, 2017; Chen et al., 2019). This work presents the results of several machine learning methods applied in geochemistry of rocks produced by volcanic eruptions at the Neapolitan area in Italy (Fig. 1).

The main benefits of using machine learning techniques are:

• Machine learning algorithms are very general and, without any specific change, except the input data, they can be applied to a huge variety of different problems despite their differences. So basically there is no need to put any efforts to find any topic-specific mathematical or statistical approach as well as specific patterns into data (in other words, machine learning methods find themselves such approaches or patterns). That's the reason why such techniques are becoming very popular in many different areas as medical diagnosis, economical choices, face authentication, scams recognition and many others;

• The only effort required is providing enough data to let the machine learn how to solve the problem and then the system will be able to autonomously and automatically solve the same problem in different situations, assuming that such situations are similar to the provided ones. Despite similar is not an objective definition, the machine learning ability to generalize the provided data can be measured after the learning;

• Many times they achieve great results even if available data are heterogeneous;

• Even if machines cannot replace humans and these algorithms have to be considered as support to scientists, machines lack of emotions can be very advantageous to avoid statistical bias. In fact, sometimes, for scientists, it can be difficult to be unbiased. For example, this is the reason why medical statistical procedures require double blinded protocols.

There are some applications in rock geochemistry (Malmgren and Nordlund, 1996; Petrelli et al., 2016, Petrelli et al., 2017; Jiao et al., 2018; Zhang et al., 2019); unsupervised and self-organizing techniques are also for the Neapolitan area (Esposito et al., 2020a, Esposito et al., 2020b). Differently from previous works, we use a huge dataset of unselected 9800 samples related to 384 eruption formations (or units) and 54 variables concerning common both whole-rock and glass chemistry. Furthermore, rather than using a single machine learning technique we applied several methods to find the one showing the best accuracy. These techniques “learn” from data input, so they should be able to unravel information even if there is data disproportion (not uniform distribution among series) and inhomogeneity (absence of certain parameters or predictors in series), and of heterogeneous composition of eruptive products and single rock units. Our aim is to explore the efficiency of machine learning approaches to automatically elaborate rock geochemistry data. More specifically, our main goal was getting an answer to the following question: “can machine learning approaches be used to predict the volcano source, the eruptive formation and the eruption period of an unknown deposit on the basis of rock geochemistry?”