Constraining the Collisional History of Asteroid (16) Psyche Using Machine Learning
Total Page:16
File Type:pdf, Size:1020Kb
51st Lunar and Planetary Science Conference (2020) 1486.pdf CONSTRAINING THE COLLISIONAL HISTORY OF ASTEROID (16) PSYCHE USING MACHINE LEARNING. S. Cambioni1, E. Asphaug1, T. S. J. Gabriel2, A. Emsenhuber1, M. Jutzi3, K. Sugiura4, S. R. Schwartz1, B. Weiss5 . 1Lunar and Planetary Laboratory, University of Arizona; 2School of Earth and Space Exploration, Arizona State University; 3Physics Institute, NCCR PlanetS, University of Bern, Switzerland; 4Tokyo Institute of Technology, Meguro, Japan. 5Massachusetts Institute of Technology, Cambridge, MA. (corresponding author: [email protected] zona.edu). Introduction: Asteroid (16) Psyche in the main as- Here we describe the possibilities emerging from the teroid belt [1] is believed to have a strong excess of application of machine learning to approach the prob- core-type materials compared to chondritic bodies. lem from both ends, making best use of the existing Does this mean it is a core that lost its mantle? If so, did “training data” (SPH simulations) in order to make pre- this mass loss happen before, during, or after differenti- dictions for observations (Figure 1). It will allow us to ation? The prevailing idea is that (16) Psyche, perhaps constrain theories by using not only bulk density but like the parent bodies of the iron meteorites, lost its also relative abundances and spin states and family mantle in a catastrophic disruption of some kind [2]. But preservation. Psyche, which may be on the boundary be- any simple answer involving collisions is challenged by tween surviving planetary embryos and disrupted al- problems. Where did the mantle fragments go? If Psy- most-planets, thereby provides a great deal of leverage che is more complex than a simple core remnant, can a in constraining planet formation theories. series of collisions explain it? Machine learning & giant impacts: In N-body Modelling the origin of Psyche from collisions has a planet formation studies, running hydrocode simula- lot in common with the problem of understanding the tions “on the fly” is impractical, unless resorting to low “late stage” of terrestrial planet formation [3], but ex- particle resolutions (e.g., [13]) or to scaling laws based tends to the more complex collisional regime involving on physical assumptions (e.g., [14]). This motivates the small bodies [e.g. 4, 5, 6], starting from Vesta or Ceres- use of machine learning to increase the realism of colli- size progenitors [7]. The systematic exploration of pos- sion models in studies of planet formation by giant im- sibilities requires a rapid way of going through a large pacts [8]. This approach allows predicting the collision parameter space, and a mechanism to apply diverse sets outcome quickly and at a known level of accuracy with of astronomical and mission data as constraints. For respect to high-resolution SPH simulations. The algo- gravity-dominated giant impacts, machine learning has rithms –“surrogate collision models”– are fully data- been applied [8, 9] to map input parameters (velocity, driven: they do not rely on model assumptions and do impact angle, mass ratio) to basic outputs like accretion not overfit. We have implemented surrogate collision efficiency, and the transition from accretion to hit-and- models in a N-body-compatible routine called collre- run. These machine-generated functions can be used as solve [9] (https://github.com/aemsenhuber/collresolve) forward models to assess the merits of hypotheses, and to mimic high-resolution hydrocode simulations during to obtain detailed information about composition and planet formation studies. thermodynamic history related to a given evolution. These surrogate models are valid for giant impacts. Unfortunately there is not yet the same kind of data- Extrapolation to the asteroid regime is not advised, for base for smaller-scale collisions, 100-1000 km, involv- the reasons mentioned. In the regime relevant to the ing Psyche and its progenitors, as there is for giant im- origin of Psyche we are training new surrogate models pacts. The general trends are the same: the projectile and to predict the collision outcome between asteroid-size target often do not merge and if they do, some amount bodies as a function of colliding masses, impact velocity of debris is often lost. But the physics of the collisions and geometry, and material properties, such as strength, are more complex, representing a transition regime porosity and friction. The outcome is described in terms where shocks, hydrostatic compressions and decom- of type of collision (e.g., merging of the two bodies or pressions can all be important, in addition to friction, their disruption), the mass of the two largest remnants, cohesion and tensile strength. Then there is the greater and the degree of impact heating and compaction. The significance to realistic thermodynamics as it influences new surrogate models will be trained on existing high- melting and rheology [10, 11, 12]. There is furthermore resolution simulations of collisions between similar- a computational problem: a binary merger takes several sized small bodies (e.g., [10; 11]). We also retain the gravitational timescales (hours to days) regardless of capability to expand the dataset using PKDGRAV (e.g., size, while the physical timestep is proportional to the [5]). As the training dataset increases, this will ulti- size (Courant criterion). mately allow us to achieve high accuracy at training and testing over the entire parameter space. 51st Lunar and Planetary Science Conference (2020) 1486.pdf Figure 1. Schematics of the envisioned approach to constrain the likelihood of specified collision scenarios of aster- oid (16) Psyche in the more general context of terrestrial planet formation. The “collision physics kernel” of the methodology is based on the Machine Learning (ML) of giant impact SPH simulations (methods in [8, 9]). The combination of ML and Bayesian (Markov Chain Monte Carlo) statistics allows for inversion of outcomes pro- vided by the surrogate model [e.g., 15]. Origin of Psyche: In the inner solar system, it is core and promote generation of a magnetic dynamo thought to be likely for Mercury to be a multiple hit-and- [22]. Repeating this study for (16) Psyche may provide run collision survivor [7, 16]. Asteroid (16) Psyche may context to interpret the measurements by the on-board have lost its silicate mantle in a similar scenario. But magnetometer and by the X-band Gravity Science In- also, the asteroid could have any number of origins and vestigation (in terms of the interior structure). be consistent with the wide range in densities and bulk Acknowledgments: S.C., A.E., and S.R.S. compositions that are presently poorly constrained (e.g., acknowledge support from NASA 80NSSC19K0817. [17, 18, 19], and references therein). In the idealized E.A. is supported by NASA NNM16AA09C Psyche: case where Psyche is a ball of metal, one can infer that Journey to a Metal World. it was stripped by repeated hit-and-run collisions. If it References: [1] Elkins-Tanton, L.T. et al. (2020). has an almost equal fraction of silicates, then a “garden Observations, meteorites, and models: A pre-flight as- variety” of hit-and-run collision would suffice. Mission sessment of the composition and formation of (16) Psy- data will resolve a lot of this uncertainty by putting con- che. JGR Planets, in review. [2] Davis, D.R. et al. straints on bulk composition and mass density, and (1999). Icarus, 137, 140. [3] Wetherill, G.W. (1985), providing information about past and present geody- Science, 228, 877-879. [4] Jutzi, M. (2015). Plan. & namic activity. As no planetary relic is the result of just Space Sci. 107, 3-9. [5] Schwartz, S.R. et al. (2018). Na- one collision, we must be aware of a range of formation ture Astronomy 2, 379. [6] Asphaug, E., (2017). Chap- ter in “Planetesimals: Early Differentiation and Conse- pathways, i.e. collision histories. quences for Planets” (Elkins-Tanton, L.T. et al., eds.), Additionally, episodes of late accretion onto terres- Cambridge University Press. [7] Asphaug, E. & Reufer, trial planets are characterized by a high degree of impact A. (2014), Nature Geosci. 7, 2014. [8] Cambioni, S. et heating, which is likely to create deep magma oceans on al. (2019), ApJ, 875, 40. [9] Emsenhuber, A. et al., the remnants [20]. We are currently working on com- (2020) arXiv preprint: 2001.00951. [10] Sugiura, K., et bining the surrogate collision models [8, 9] with the sil- al. (2018). Astronomy & Astrophysics, 620, p.A167. icate-metal equilibration codes by [21] to study the ef- [11] Jutzi, M., (2019). Planetary and Space Science, fect of inefficient accretion on core formation and 177, p.104695. [12] Asphaug, E. et al. (2017), Global chemical element partitioning as the planets accretes. Scale Collisions, in Asteroids IV. [13] Burger, C., This methodology can be applied to understand how as- (2019). arXiv preprint: 1910.14334. [14] Leinhardt, teroid (16) Psyche differentiated. This will provide test- Z.M. & Stewart, S.T. (2011) Astrophys. J., 745, p.79. able hypotheses with measurables prior to the NASA’s [15] Cambioni, S., et al. (2019), Icarus, 325, pp.16-30. Psyche mission’s arrival at the asteroid. The silicate- [16] Chambers, J.E., (2013). Icarus, 224(1), pp.43-56. metal equilibration model provides predictions of the [17] Elkins-Tanton, L.T. (2018) Elements, 14.1: 68-68. amount of metal and silicate constituents in the core. [18] Viikinkoski, M. et al. (2018). Astronomy & Astro- This is relevant to measurements by the NASA’s Psyche physics 619: L3. [19] Siltala, L. and Granvik, M., Multispectral Imager and Gamma-ray and Neutron (2019), EPSC, Vol. 13, 1610-1. [20] Tonks, W.B., & Spectrometer. Finally, energetic collisions involving Melosh, H.J. (1993). JGR: Planets, 98, 5319-5333. [21] terrestrial planets have been found to be able to mix the Rubie, D.C., et al.