Home      Publications       Courses      Research

____________________________________________________________________________

 

Inverse Problems in Geodynamics Using Machine Learning Algorithms

 

During the past few decades numerical studies have been widely employed to explore the style of circulation and mixing in the mantle of Earth and other planets. However, in geodynamical studies there are many properties from mineral physics, geochemistry and petrology in these numerical models. Machine learning (ML), as a computational statistic-related technique and a subfield of artificial intelligence (AI), has rapidly emerged recently in many fields of sciences and engineering. We focus here on the application of supervised machine learning (SML) algorithms in predictions of mantle flow processes. Specifically we emphasize on estimating mantle properties by employing machine learning techniques in solving an inverse problem. Using snapshots of numerical convection models as training samples we enable machine learning models to determine the magnitude of the spin transition-induced density anomalies that can cause flow stagnation at mid-mantle depths. Employing support vector machine (SVM) algorithms we show that SML techniques can successfully predict the magnitude of mantle density anomalies and can also be used in characterizing mantle flow patterns. The technique can be extended to more complex geodynamics problems in mantle dynamics by employing deep learning algorithms for putting constraints on properties such as viscosity, elastic parameters, and the nature of thermal and chemical anomalies.

 

Description: E:\BP\Papers\Paper-23_Big_Data\Submission01_JGR_N\submitted2\Figures\Fig01\Fig01.JPG      Description: E:\BP\Papers\Paper-23_Big_Data\Submission01_JGR_N\submitted2\Figures\Fig02\Fig02.JPG

Figure 1 – Artificial neutron of McCulloch and Pitts.

Figure 2 – a) Temperature fields after 200 Myr evolution for the models with the density anomalies represented by m = 0, m = 100, m = 200, and m = 299 b) Close-up of  a temperature filed showing the stagnated slabs and plumes as described in the text.

Description: E:\BP\Papers\Paper-23_Big_Data\Submission01_JGR_N\submitted2\Figures\Fig03\Fig03.JPG

Figure 3 – Grid search results for the first inverse problem (M300 samples) a) Coarse grid search results using SVC-Linear, b) Fine grid search histogram for the model specified in (a), c) Coarse grid search results using SVC-Poly, d) Fine grid search histogram for the model specified in (c), e) Coarse grid search results using SVC-RBF, f) Fine grid search histogram for the model specified in (e). 

 

Description: E:\BP\Papers\Paper-23_Big_Data\Submission01_JGR_N\submitted2\Figures\Fig04\Fig04.JPG

Figure 4 – Inverse problem learning models using M300 samples: a) SVC-Linear b) SVC-Poly model, c) SVC-RBF model. The training samples are shown as scattered plot (black circles). The fill-color of the circles mismatches with the background class color in the misclassified samples.  The slab and plume volumes are normalized to 1. The prediction accuracies for the linear, polynomial degree 2 and RBF-kernel are about 91%, 91% and 92% respectively. For each pair of sample features, the class color identifies the class label with respect to the legend.

 

 

Alborz: A Parallelized Spherical Control Volume Convection Code (PSCVCC) (developed by H. Shahnas)

 

A Three-Dimensional Control-Volume based Model of Planetary Interior Convection

 

 

We employ a control volume methodology to develop a new anelastically compressible model of three-dimensional thermal convection in the "mantle" of a terrestrial planet that fully incorporates the influence of large variations in material properties, both in radius and in azimuth. Parallelization of the software employs the Message Passing Interface (MPI), and the code has been ported to a Power 6 cluster and bench-marked against an existing model which is not able to accommodate very large lateral variations of material properties. Results for models characterized by different Rayleigh numbers and rates of internal heating and constant physical properties demonstrate that the control volume results are consistent with previously published results produced by TERRA, a finite element code which employs an icosahedral grid.  Compared with this and other formulations of the problem of high Rayleigh number convective mixing, the control volume formulation is demonstrably superior. The numerical model thoroughly documented herein also incorporates the influence of solid-solid pressure-induced phase transformations, as well as arbitrarily pressure and temperature dependent physical properties.  When applied to the mantle of the Earth, models which include the exothermic and endothermic phase transitions at 410 km and 660 km depth respectively, and a radial viscosity profile derived from the inversion of glacial isostatic adjustment data, exhibit similar radial layering of the flow as previously shown to be characteristic of  axi-symmetric spherical models. This layering, an extremely important aspect of the style of convective mixing insofar as the understanding of Earth evolution is concerned, is shown to be relatively unaffected in the three dimensional model by the action of the recently discovered additional pressure induced phase transformation that occurs just above the boundary between the solid mantle and the liquid iron outer core.

 

a)     Spherical volume element, b) latitude-longitude grid employed, c) velocity field at a constant depth due to a Gaussian temperature anomaly at CMB which is sinusoidally vanishing at the top surface, and d) slice subdomains which communicate through the MPI.

b)      

 

The second version of the code uses Yin-Yang grid,

 

 

Description: C:\D\000\Tetra_7E3.jpg

 

 

Description: C:\D\Work_Shop_D\Paper-14\Version-03\Figures\Fig02\Fig02_Web.JPG

 

Temperature snapshots of the models C00D00VM3 and C12D00VM3. The mantle plumes stagnated at the 660 km depth phase boundary horizon, are clearly detected in both panels in the first column. The mantle avalanches are also detectable from the panels of the second column. Superimposed slices of line contours provides a better visualization of the layered convection.

 

Left: A) Temperature contours at a depth about 2370 km for the model Ra1E6-VT with constant physical properties and depth and pressure-dependent viscosity of the form described by equation (60-1) with constant radial profile , B) Viscosity contours corresponding to the temperature contours in (A) with the velocity arrows superimposed on this surface , C) Temperature isosurfaces demonstrating the style of convection.

 

Middle: A) Temperature contours for the model Ra1E6-VT at about 150 km depth, B) The corresponding viscosity contours demonstrating a lateral viscosity contrast of magnitude  ~3700.

 

Right: A) Temperature contours at a depth about 2370 km for the model  C12D00VM3 with the upper mantle exothermic and endothermic and deep mantle exothermic phase transitions and also depth and pressure-dependent viscosity of the form described by equation (60-1) with VM3 radial profile , B) Viscosity contours corresponding to the temperature contours in (A) with the velocity arrows superimposed on this surface , C) Temperature isosurfaces demonstrating the style of convection.

 

 

 

Animations of Mantle Convection

 

Avalanches1    

Avalanches2

Avalanches3

3D-C-Volume1

3D-C-Volume2

3D-C-Volume3

 

 

 

 

Convection in a Spherical Shell with Basal and Internal Heating:

 Implications for the Thermal State of the Thin-lithosphere Single Plate Planets

 

The parallelized spherical control volume convection code (PSCVCC) is used to study the planforms of convecting mantle and a scaling relationship between the mean mantle temperature, the rate of internal heating and the Rayleigh number for the planets in which there is no significant surface tectonics in three-dimensional spherical geometry. Similarly we present a power law relation for the surface heat flux in terms of the Rayleigh number and mean mantle temperature. In a systematic study we employ a large number of 3D-models characterized by different Rayleigh numbers and rates of internal heating and constant physical properties to investigate the pattern of the convection and the thermal characteristics of the convecting layer. Inversion of the numeric results for the non-dimensional mean mantle temperature and surface heat flux yields to the simple power law relations specifying the efficiency of the heat transfer in the planet. The parameterized power law relation is then extended to a wide range of curvatures parameter f. Compared to the Earth-like planets in the single-plate planets with thin lithosphere under similar conditions (the same Rayleigh number and shell curvature), the mantle is higher in temperature due to the non-convecting lithosphere layer in the latter case. However, due to the reduction in the vigor of near surface convection there is a reduction in the surface heat flux in the single plate planets.

 

 

Temperature field variation as a function of the Rayleigh number Ra and internal heating rate for Earth-like curvature (). The left panel at each snapshot demonstrates two cold and hot isosurfaces with temperatures  and  respectively where  is the mantle mean temperature. The right panel shows a cross-section of the mantle convection. A) Snapshots for the models with no internal heat sources, B) Snapshots for the models with non-dimensional internal heating rate of 5.4, C) Snapshots for the models with non-dimensional internal heating rate of 10.79, D) Snapshots for the models with non-dimensional internal heating rate of 16.19.

 

 

Temperature field variation as a function of the Rayleigh number Ra and internal heating rate for a number of models with curvature parameters between 0.1 and 0.9. The left panel at each snapshot demonstrates two cold and hot isosurfaces with temperatures  and  respectively where  is the mantle mean temperature. The right panel shows a cross-section of the mantle convection. . A) Snapshots for the models with and rate of internal heating of 27, B) Snapshots for the models with and rate of internal heating of 27, C) Snapshots for the models with and rate of internal heating of 16.2, D) Snapshots for the models with and rate of internal heating of 16.2, E) Snapshots for the models with and rate of internal heating of 5.4, F) Snapshots for the models with and rate of internal heating of 10.8 (the internal heating rates and temperatures are in non-dimensional form).

 

 

 

 

The High Pressure Electronic Spin Transition in Iron:

 Potential Impacts upon Mantle Mixing

 

At the Rayleigh number appropriate to Earthfs mantle, radial heat transport is dominated by solid state thermal convection. Because of the large number of physical properties required to determine the Rayleigh number, and because these properties are expected to be (perhaps strong) functions of pressure and temperature (P-T), laboratory measurements of them under the high pressure and temperature conditions that occur in the deep Earth are of fundamental importance. Recent experimental data demonstrate that an electronic spin transition in iron that occurs at midmantle depths results in significant changes in the physical properties of the ferropericlase component of mantle mineralogy. Additional recent results suggest that it may also exist in the dominant perovskite component. Using control volume based numerical models we investigate the impacts on mantle mixing of this spin transition through its influence on the most important subset of these physical properties, namely density, thermal expansivity, bulk modulus and heat capacity. Our numerical model results demonstrate that this electronic transition enhances mixing in the lower regions of the lower mantle by enhancing the vigor of rising plumes. The lowermost region of the mantle is slightly warmed and the upper mantle slightly cooled by spin]induced effects. However, the spin crossover in the lower mantle appears not to significantly influence mantle layering. Due to the competition that could exist between the strength of the spin-induced thermodynamic properties of ferropericlase and perovskite, cold descending thermal anomalies could stagnate at middle-to-lower mantle depths and lead to the occurrence of ”mid mantle avalanches”.

 

 

Snapshots of the spin-induced mid-mantle avalanche (SIMMA) event in the model  (CMB at 4000 K) described in section 4.4 in 10 Myr intervals. The first and second images at each snapshot show the temperature and spin-induced anomaly of  form which the laterally averaged anomalies have been subtracted at each grid point (i.e. ) respectively.

 

 

 

 

 

 

3D Convection and the Geometrical Effects of Curvature in the Efficiency of Heat Transfer in an Isoviscous

Fluid Heated from Within and Below at High Prandtl Number

 

Thermal cooling models are essential in understanding the thermal history of terrestrial planets. The thermal history of the cooling Earth have been investigated by many authors during the last few decades (Tozer, 1967; McKenzie & Weiss, 1975, 1980; Shrpe & Peltier, 1978, 1979; Schubert, Cassen & Young, 1979; Turcotte, Cooke & Willeman, 1979; Davies, 1980; Peltier & Jarvis, 1982; Richter, 1984). Cooling of the interior of terrestrial planets and how the interior temperatures are related to surface heat flux requires extensive study of convecting heat flow in real geometry.  During the last three decades intensive efforts have been made to understand the different aspects of heat transfer and the effect of controlling parameters in mantle dynamics employing two dimensional numeric models (Vangelov and Jarvis, 1994; Jarvis, 1993; Jarvis et al., 1995). However the actual geometry of mantle convection is three dimensional with spherical curvature at the top and bottom boundaries. During the last decade due to improved computer and software efficiency, large 3D numerical calculations have become possible and 3D numeric models have been employed in the field of Earth sciences and particularly in mantle convection. Parallel computing techniques and more efficient algorithms for large problems in computational science have now made many numeric studies possible. Some works have been done on Cartesian and spherical geometries. The effect of the Rayleigh number and aspect ratio on the planform and heat transfer has been studied by Travis et al. (1990) and Sotin and Labrosse (1999). 

In this study a systematic investigation is made to explore the effect of the curvature on the cooling of the planets using isoviscous models heated from within and below at infinite Pandtl number.  TERRA  a 3-D spherical finite element mantle dynamics code (Baumgardner in 1983; parallized by Bunge in 1993; Bunge and Baumgardner, 1995) has been used.

 

 

 

 

Description: 204-11Description: 213-3Description: 220-16

(a)                        (b)                         (c)

 

 

Models with different internal heating rates and Rayleigh numbers.

 

 

 

 

 

 

The Effect of Mechanical Boundary conditions on the Secular Cooling of the Planetary Mantles

 

It has been suggested that the lack of an appreciable Venusian dipolar magnetic field has resulted from the absence of lithospheric participation in Venusian mantle convection for the past 0.5-0.75 billion years (during this period mantle convection has been limited to the thick stagnant-lid regime). In contrast to convection in the stagnant-lid regime, terrestrial mantle convection (featuring the subduction of oceanic lithosphere) is particularly efficient at cooling the Earth. Moreover, participation of the Earth's outer thermal boundary layer in terrestrial mantle convection carries cool material deep into the planetary interior and influences heat flow at the core-mantle boundary. Thus, lithospheric subduction influences the geodynamo that originates with compositional convection in the conducting outer core.

 

 

Description: model-353-step30-Free-WithCoreCooling-highTDescription: model-351-step30-Rigid-WithCoreCooling-highT

Figure 1 -Temperatures for depth-dependent  viscosity models with free surface. Green isosurface corresponds to temperature of 1493 K and orange surface corresponds to a temperature of 2393 K.                                                                                                 

Figure 2 -Temperatures for depth-dependent viscosity models with rigid surface (initially free-slip surface). Green isosurface corresponds to temperature of 1493 K and orange surface corresponds to a temperature of 2393 K. Reduction in heat flow out of the core of up to 50% during the period that elapses between the onset of surface immobility and the final state reached in this model which may weaken or halt thermal convection in the core and hence the generation of the global magnetic field.

                                                                                                                                                              

 

 

Layered Intra-lithospheric Small-Scale Convection in the Ionian Asthenosphere:

Implications for Short-Range Surface Topography and Heat Flow

 

Io, one of the four Galilean moons of Jupiter is remarkable for its extensive volcanism and extremely low viscosity asthenosphere. The implied degree of convective destabilization of the asthenosphere is characterized by an extremely high Rayleigh number of O(1012), suggesting vigorous thermal convection in addition to the significant internal heat generation by tidal friction that must be transported from the interior to the base of the Io lithosphere. This radially advected heat is evacuated to the surface by both conduction and volcanism. Despite Io’s ubiquitous volcanism only 4% of its mountains (montes) appear to have a volcanic origin and most of the mountains have been initiated by tectonic processes. By employing control volume numerical models we have investigated the style of convection in the interior of Io and its correlation with the Ionian surface heat flux and topography. Our control volume results support the occurrence of significant asthenosphere heating and demonstrate that the short wavelength features of the surface heat flux are well correlated to an expected layered intra-lithospheric small-scale convection (LILSSC) style. These analyses suggest that the amplitude of the short wavelength topography of Io is expected to be an order of a few hundred meters. The model results also demonstrate that the Ionian highs cannot be produced by a lithospheric flexure process above the hot upwellings and therefore other tectonic events, as have previously suggested, must be responsible for the formation of the high Ionian mountains that reach in excess of 17 km in elevation. 

 

(a) Snapshot of the temperature field (K) with superimposed velocity arrows at its statistically steady state, (b) the logarithm of the velocity field (m/yr) with the superimposed velocity arrows.

 

 

 

Viscous and Impact Demagnetization of the Martian Crust

 

The magnetization of the Martian crust has been modified since the cessation of the core dynamo at around 4 Gyr ago, partly by impact-induced shock waves at shallow depths and partly by viscous decay at deeper parts.  The thermal evolution models of Mars suggest that the potentially magnetic layer was about 85 km thick during the active period of the core dynamo, assuming magnetite as the major magnetic carrier. The viscous relaxation of the major magnetic carriers suggests that the depth to the lower boundary of the magnetic layer has gradually decreased, by a total of about 30 km, in the last 4 Gyr if magnetite is the major magnetic carrier.   The large impacts that created the giant basins on Mars, Hellas, Argyre and Isidis, have almost completely demagnetized the crust beneath the basins. On the basis of the shock wave pressures produced by smaller impacts that created craters of diameters 300 – 1000 km it is expected that the crust would be significantly demagnetized beneath these craters. However, except for a few craters, there is no signature of appreciable demagnetization.  This emphasizes that either the magnetic carriers have high coercivity and resisted demagnetization, or magnetic source bodies are deep seated, or they have acquired magnetization after the intensive impact cratering period.  An alternate possibility is that the scaling laws proposed for very small craters do not adequately apply to the large craters considered in this paper.

 

Description: Fig8A_rev-2Description: Fig8B_rev-2Description: Fig8C_rev2

(a)                        (b)                            (c)

 

Figure 1.  The normalized magnetization (normalized to 1 at 4 Gyr ago) of the magnetic layer as a function of time, for magnetite (a), hematite (b) and pyrrhotite (c). Panels (d) and (e) show the thermal (with no viscous decay) and viscous (only) demagnetization, respectively. Panel (f) shows the thermal evolution within the upper 100 km. The I-series are the nominal stagnant lithosphere model.  The II and III-series are for plate tectonics models operating in the first 250 Myr and 500 Myr, respectively.

 

 

 

 

 

 

Anomalous topography in the western Atlantic caused by edge-driven convection

 

The western Atlantic region contains a long wavelength intraplate topography anomaly that is defined by the NE-SW trending Bermuda Rise and two adjacent topography lows. Numerical models are used to investigate the hypothesis that the anomalous topography may be the surface response to edge-driven convection.

A primary edge-driven convection cell and secondary flow circulation develops at a modeled continent-ocean plate margin and induces subsidence at the continent-ocean margin, an offshore peak/plateau of high topography on the ocean plate, and distal ocean plate subsidence. Unlike hot spots, the edge-driven convection cell and associated topography migrate with moving surface plates. The flow cell and wavelength of topography is broadened with continent-ward motion of the lithosphere relative to the mantle, whereas a migration in the ocean-ward direction suppresses the formation of the edge-driven convection cell and surface topography. The wavelength of observed anomalous topography in the western Atlantic and estimates of plate motions relative to a fixed hot spot reference frame are consistent with the former.

Description: image005

Figure 1. (a) Bathymetric map of the western North Atlantic region. (b) Profiles of residual topography across the sections shown in Figure 1a and an averaged section (bold solid line). The residual topography is surface elevation corrected for isostatic effects (crustal thickening on the continental plate and lithospheric cooling and subsidence on the ocean plate) and sediment loading. These corrections have been applied separately for continental plate and ocean plate to derive the profiles shown; see text for detailed explanation. Arrow marks the approximate location of the ocean-continent plate transition.

 

Description: image008Description: image009

Figure 2. Temperature fields and flow velocity vectors at (a) 22 and (b) 47 Myr. (c) Profiles of surface topography at time intervals as indicated. Abscissa distances are from the left side of the solution space; arrows mark the location of the continent-ocean boundary.

 

 

On the Relative Importance of Mineral Phase Transitions and Viscosity Stratification in

Controlling the Sinking Rates of Detached Slab Remnants

 

It has been known for some time that, in one respect, the effect on mantle convection of an endothermic phase  transition of olivine at 660 km depth is qualitatively similar to the effect of a chemical or viscosity stratification with higher density, or viscosity, in the lower mantle (e.g., Richter, 1973; Davies, 1977;  Christensen and Yuen, 1985).  In each case, the ability of convective flow in the upper mantle to penetrate  into the lower mantle is either inhibited or prevented, depending on the degree of stratification.  Consequently, when considering a dynamic explanation for the mid-mantle thermal anomalies below India and Tibet, as described by Van der Voo et al. (1999),  Jarvis and Lowman (2005) invoked the effects of viscosity stratification as a possibility.  However, for viscosity increases up to a factor of 65 across the mantle, these authors could not account for the shallow depth of the most northern of the cold thermal anomalies.  Van der Voo et. (1999) had interpreted this anomaly as the remnant of a cold slab from a fossil subduction zone which was terminated 141 Myr ago (Besse and Courtillot, 1988). Subsequently, Jarvis and Lowman (2006) modeled the sinking of three-dimensional slab remnants in the presence of a step-increase in viscosity at 660 km.  They found that slab remnants from short-lived subduction zones (which had not penetrated into the deep mantle before subduction ended)  were retained in the upper mantle for times much greater than 150 Myr, while larger remnants from mature subduction zones (which had extended across the whole mantle) sank into the lower half of the mantle in less than 150 Myr.

 

Since the northernmost anomaly under India/Tibet is believed to be a 141 Myr remnant of a mature subduction zone,  having ingested some 4,000 km of the Paleo-Tethys sea-floor, we cannot attribute its persistence at a relatively  shallow depth to short-lived prior subduction.  The presence of this enigmatic mid-mantle thermal anomaly partially motivates the present study.  We have developed a 2D cylindrical-shell control-volume numerical model which includes an exothermic phase transition at 410 km depth, an endothermic phase transition at 660 km and depth-dependent viscosity  stratification. Our numerical code carries an associated Lagrangian grid which is initially redundant  with the Eulerian grid.  When activated it may be used for fluid particle tracking and/or as a material  property carrier. We study the time taken for cold parcels of mantle material to sink from the upper  surface to the lower surface in models with and without mineral phase transitions and for various degrees of viscosity stratification.  Here, the relative importance of phase transitions and viscosity  stratification is assessed.

 

Description: Fig1Description: Fig2  

Figure 1. (a) Isotherms at an arbitrary time, designated T = 0 Myr, in a constant viscosity model  (Model C-1) with an endothermic phase transition at 660 km depth. (b) Model C-1 evolved  a further 76 Myr. The arrow shows the location of the first old parcels from the upper surface to reach the CMB since T = 0 Myr.  (c) Highlighted rows  on the Lagrangian grid at T = 0 Myr. (d) Corresponding evolution of the highlighted rows on the Langrangian grid after 76 Myr. The contours are equally spaced over the temperature range with the steps of 67 K.                  

Figure 2.  (a) Isotherms of a constant viscosity model with no phase transitions , Model C-0, at the time (40 Myr) when the first cold parcels from the upper surface arrive at the CMB since T = 0.  The solid arrows show the location of the arriving material; the dashed arrow indicates the location of material which arrives at t = 45 Myr. (b) Isotherms of a constant viscosity model with two phase transitions , Model C-2, at 50 Myr when the first cold parcels from the upper surface arrive at the CMB since T = 0. The dashed arrow shows the location of the arriving material.  The two solid arrows indicate two sites of incipient avalanches. (c) Isotherms of a stratified viscosity model with no phase transitions , Model S65-0, at 70 Myr when the first old parcels from the upper surface arrive at the CMB since T = 0. The arrows show the location of the arriving material. (d) Isotherms of a stratified viscosity model with two phase transitions , Model S65-2, at 70 Myr when the first cold parcel from the upper surface arrives at the CMB since T = 0. The arrow shows the location of the arriving material. The contours are equally spaced over the temperature range with the steps of 67 K.

 

Description: C:\Shahnas\000Conf\HTMLS\index_files\image028.png

Figure  3.  (a) Stratified (step) viscosity profiles employed in depth-dependent viscosity models. (b)Sinking times for different upper and lower mantle viscosity contrasts in stratified viscosity models.