ORALS

SESSION: GeochemistryTuePM2-R8	Navrotsky International Symposium (2nd Intl. Symp. on Geochemistry for Sustainable Development)
Tue. 28 Nov. 2023 / Room: Coral Reef
Session Chairs: Qijun Hong; Fabienne Trolard; Session Monitor: TBA

16:25: [GeochemistryTuePM210] OS
DEEP LEARNING FOR LARGE-SCALE PREDICTION OF MELTING TEMPERATURE AND MATERIALS PROPERTIES
Qijun Hong¹ ; ¹Arizona State University, Tempe, United States;
Paper Id: 319 [Abstract]

High-temperature materials properties are challenging to compute from first principles or machine learning. We demonstrate the feasibility of rapid and large-scale prediction of melting temperature prediction based on integrated density functional theory calculations and deep learning. We employ the method and tool to discover the material with the highest melting temperature in the world, as well as dozens of potential candidates of refractory materials.We build cyber infrastructure to provide public access to our databases and models. Our application programming interface (API) enables users to swiftly calculate the melting temperatures of as many as 10,000 materials in a single API call, with a processing speed of 0.03 seconds per material. This valuable functionality is accessible to the general public free of charge, with all calculations being handled by our server hosted at the ASU Research Computing Center. An alternative way for users to access our models is by visiting our website through a browser, where they can perform single material calculations.In order to extend the applicability of our methodology, we have constructed deep learning models for the prediction of various material properties beyond melting temperature, such as bulk modulus, volume, fusion enthalpy, and others. We build the Materials Properties Prediction (MAPP) framework, characterized by a diverse array of materials properties, the potential for iterative enhancement, and the prospect of model integration for systematic improvement.

References:
[1] Hong, Q.-J, S.V. Ushakov, A. Navrotsky, and A. van de Walle, "Melting temperature prediction using a graph neural network model: from ancient minerals to new materials,", PNAS, 119.36 (2022): e2209630119.
[2] Hong, Q.-J, "Melting temperature prediction via first principles and deep learning,", Computational Materials Science, 214 (2022): 111684.
[3] Hong, Q.-J, A. van de Walle, S.V. Ushakov, and A. Navrotsky, "Integrating computational and experimental thermodynamics of refractory materials at high temperature,", CALPHAD, 79 (2022): 102500.

SESSION: GeochemistryTuePM2-R8	Navrotsky International Symposium (2nd Intl. Symp. on Geochemistry for Sustainable Development)
Tue. 28 Nov. 2023 / Room: Coral Reef
Session Chairs: Qijun Hong; Fabienne Trolard; Session Monitor: TBA

17:15: [GeochemistryTuePM212] OS
HIGH TEMPERATURE THERMOCHEMISTRY FROM EXPERIMENT, AB INITIO, AND MACHINE LEARNING
Sergey Ushakov¹ ; Qijun Hong¹ ; Alexandra Navrotsky² ; ¹Arizona State University, Tempe, United States; ²Arizona State University, Phoenix, United States;
Paper Id: 14 [Abstract]

The measurements, computations, and predictions of high temperature thermodynamic properties are of interest to geoscience, material science, and engineering. The experimental techniques to provide structural and thermodynamic data above 1500 °C were developed in Navrotsky’s group for over 10 years. This resulted in the first demonstration of crystal structure refinements on laser-heated aerodynamically levitated samples using synchrotron X-ray and neutron diffraction and drop calorimetry measurements with splittable nozzle aerodynamic levitator [1]. High temperature diffraction provides experimental data on thermal expansion, atomic displacement parameters, and volume change in phase transformations. Drop calorimetry on levitated samples provides enthalpy of fusion. These data can also be obtained from ab initio molecular dynamic computations. The experimentally benchmarked computations can provide reliable data on high temperature heat capacities [2]. Melting or decomposition temperature is a widely used thermodynamic property. Experimental measurements and ab initio computations require time, resources, and expertise. The machine learning model has been developed and trained on ~10,000 experimental and ab initio values of melting points for congruently melting compounds. It has been applied to predict melting or decomposition temperatures of ~5,000 known mineral species which revealed new correlations with the time of Late Heavy Bombardment event and structural complexity index [3]. The model is publicly accessible via the web interface on Hong’s group website for the prediction of melting temperatures within seconds [4].

References:
[1] S. V. Ushakov, P. S. Maram, D. Kapush, A. J. Pavlik, III, M. Fyhrie, L. C. Gallington, C. J. Benmore, R. Weber, J. C. Neuefeind, J. W. McMurray, A. Navrotsky, Adv. Appl. Ceram. 117, s82-s89 (2018) https://doi.org/10.1080/17436753.2018.1516267. [2] Q.-J. Hong, A. van de Walle, S. V. Ushakov, A. Navrotsky, Calphad 79, 102500 (2022) https:/doi.org/10.1016/j.calphad.2022.102500. [3] Q.-J. Hong, S. V. Ushakov, A. van de Walle, A. Navrotsky, PNAS 119, e2209630119 (2022) https://doi.org/10.1073/pnas.2209630119. [4] https://faculty.engineering.asu.edu/hong/melting-temperature-predictor/