EGU2020-19028
https://doi.org/10.5194/egusphere-egu2020-19028
EGU General Assembly 2020
© Author(s) 2020. This work is distributed under
the Creative Commons Attribution 4.0 License.

Unravelling pre-eruptive P-T conditions by machine learning

Maurizio Petrelli1, Luca Caricchi2, and Diego Perugini1
Maurizio Petrelli et al.
  • 1University of Perugia, Department of Physics and Geology, Perugia, Italy (maurizio.petrelli@unipg.it)
  • 2University of Geneva, Department of Earth Sciences, Geneva, Switzerland

Clinopyroxene based thermometers and barometers are widely used tools for estimating temperature and pressure conditions under which magmas are stored before eruptions.

Several studies reported the development and the application of Clinopyroxene–liquid geothermobarometers in many different volcanic environments, also warning on the potential pitfall in using overly complex models [e.g., 1 and references therein]. The main drawback in the use of models with a large number of parameters is the potential overfitting of the calibration data, yielding a poor accuracy in real-world applications. On the other hand, simpler models cannot account for the complexity of natural magmatic systems, requiring different calibrations for different magma chemistries [e.g., 2, 3].

In the present study, we report on the development of Clinopyroxene and Clinopyroxene-liquid thermometers and barometers in a wide range of P-T-X conditions using Machine Learning (ML) algorithms. To avoid overfitting and demonstrate the robustness of the different methods, we randomly split the dataset into training and validation portions and repeating this procedure up to 10000 times to trace the performance of each of the used algorithms. We compared the performance of ML algorithms with classical and established Clinopyroxene and Clinopyroxene-liquid thermometers and barometers using local and global calibrations. Finally, we applied the obtained thermometers and barometers to real study cases.

 

[1]      K. D. Putirka, Thermometers and barometers for volcanic systems, Minerals, Inclusions and Volcanic Processes, 69. 61–120, 2008.

[2]      D. A. Neave, K. D. Putirka, Am. Mineral., 2017, DOI:10.2138/am-2017-5968.

[3]      M. Masotta, S. Mollo, C. Freda, M. Gaeta, G. Moore, Contrib. to Mineral. Petrol., 2013, DOI:10.1007/s00410-013-0927-9.

How to cite: Petrelli, M., Caricchi, L., and Perugini, D.: Unravelling pre-eruptive P-T conditions by machine learning, EGU General Assembly 2020, Online, 4–8 May 2020, EGU2020-19028, https://doi.org/10.5194/egusphere-egu2020-19028, 2020

Comments on the presentation

AC: Author Comment | CC: Community Comment | Report abuse

Presentation version 1 – uploaded on 04 May 2020
  • CC1: Comment on EGU2020-19028, Tonin Bechon, 05 May 2020

    Hi Maurizio,

    I have few questions about your presentation material (I prefer to avoid the chat because it is difficult to use and communicate) :

    - Is it published and is there a software that allows us to use this machine learning algorithms ?

    -Does the presence of other phases have an impact on the result ?

    -Do we have to make any assumptions on the volatile content and/or the fO2 ?

    -Instead of using 2 phases (melt and cpx) isn't it better to use all accountable ones to better constrain the system ?

    Last, i'm very happy to finaly meet the work of someone working on that topic and can't wait the mature results of such a technology.

    • CC2: Reply to CC1, Tonin Bechon, 05 May 2020

      Tonin Bechon (sorry, I forgot to sign)

    • AC1: Reply to CC1, Maurizio Petrelli, 05 May 2020

      Hi Tonin,

      - Is it published and is there a software that allows us to use this machine learning algorithms ?

      It has been just submitted to JGR. We used python sripts that we will share with the MS. But we can do some test togheter if you are interested.

       

      -Does the presence of other phases have an impact on the result ?

      We do not assume that cpx is the only phase. But we could start thinking in developing multiphase models.

       

      -Do we have to make any assumptions on the volatile content and/or the fO2 ?

      Any assumptions, it depends on the conditions of the experiments used for the calibration. We could discuss about this point.

      -Instead of using 2 phases (melt and cpx) isn't it better to use all accountable ones to better constrain the system ?

      Yes, I agree. It is the next step, in my plans.

       

      • CC3: Reply to AC1, Tonin Bechon, 05 May 2020

        Hi Maurizio,

         

        Thanks for the invitation, my topic is on a specific target () and i used Putirka and Neave and Putirka database to determine the Pressure and temperatures. As you used it for the calibration of your AI, i expect to get the same results. Concerning collaboration with you as a side project, i'll be very glad however my skills in AI or python are limited (i use matlab at the moment) and i'm not shure to be of great help for you. I registered to the workshop on machine learning next week to dig further the topic because i truly think that this technology (while dangerous in some of its aspects) may lead to great breakthrough especially in our field where geochemical dataset are too wide and we just understand part of the available informations...

        Tonin B.

        • AC2: Reply to CC3, Maurizio Petrelli, 05 May 2020

          you use the Neave and Putirka (2017) calibration ... are you working on Iceland? On the manuscript we just submitted, we made a comparison with Neave and Putirka (2017) in icelandic magmas, getting comparable results.

          • CC4: Reply to AC2, Tonin Bechon, 05 May 2020

            No I'm working on southern Chile

          • CC5: Reply to AC2, Tonin Bechon, 05 May 2020

            Neave and putirka only works for a limited measurments in my dataset du to temperature conditions (i don't have a lot of cpx at high temperature hence the limited range of my pressures i presented on my poster

          • CC6: Reply to AC2, Tonin Bechon, 05 May 2020

            Sorry i just figured out that the link wasn't accepted by the comment session my poster is in the 14h chat session