生成模型的目的是学习数据之间的复杂关系以创建新的模拟数据,但是当前的方法在高维情形下不可行。当数据的生成基于物理过程时,通过学习物理过程背后的物理规律,可以创建生成模型,获得物理过程中的对称性和约束,从而允许模型在高维情形下有较好的效果。近日发表于 PNAS 的一项研究,作者提出了拉格朗日深度学习(Lagrangian deep learning),并将其应用于学习宇宙学流体动力学模拟。
Biwei Dai、Uros Seljak | 作者
潘佳栋 | 编译
邓一雪 | 编辑
论文题目:
Learning effective physical laws for generating cosmological hydrodynamics with Lagrangian deep learning
论文地址:
https://www.pnas.org/content/118/16/e2020324118
宇宙大尺度结构形成的数值模拟对于从观测中提取宇宙信息至关重要。[1-7]原则上,流体动力学模拟能够预测宇宙中所有可观测物质的分布,从而可以模拟观测。但是,由于计算成本高,高分辨分子动力学模拟尚不可行。目前最广泛使用的方法是进行仅存在重力的N体模拟,然后使用半解析方法填充重子,这种方法有很强的假设性。[8][9]此外,许多宇宙学观测(例如X射线发射和Sunyaev–Zeldovich(SZ)发射)都是基于流体动力学的气体特性,例如气体密度、温度、压强等,在只有暗物质的情形下无法被建模模拟。 深度学习方法提供了一种替代方法,可以对宇宙可观测物质进行模拟。许多论文认为该任务是图像到图像的翻译问题,即将像素化的物质密度场作为输入数据,输出像素化的可观测场。这些方法要么使用诸如生成对抗网络(GAN)[10]和变分自编码器(VAE)[11][12]之类的深度生成模型求解条件概率分布p (ytarget|xinput),要么使用深度卷积神经网络(DCNN)学习映射xinput→ytarget。该领域以前的工作十分广泛,例如识别光晕(质子晕)[13-16],产生三维(3D)星系分布[17],生成热SZ(tSZ)信号[18],预测暗物质消亡的反馈信息[19],学习中微子效应[20]以及从低分辨率模拟中模拟高分辨率特征[21][22]等。 与这些在像素(欧拉)空间中工作并将视场视为图像的方法不同,动力学建模的另一种方法是拉格朗日方法,即通过对单个粒子或流体元素的位移场进行建模来跟踪其运动。位移场比密度场包含更多的信息,因为不同的位移场可以产生相同的密度场,并且通常比密度场具有更多的高斯分布和线性分布。该空间中的现有方法仅能求解暗物质,例如近似的N体求解器[23][24]和DCNN[25]。 在这项工作中,我们提出了一种深度学习架构,即拉格朗日深度学习(Lagrangian deep learning),使用拉格朗日方法对宇宙暗物质和流体动力学进行建模。该模型是由物理学中有效理论思想促动的,一个真实物理过程可能由于过于复杂而不能模拟,以至于无法对它进行有效的,通常是粗粒度的物理学描述。一个典型的例子是有效场论,其中对微扰场论补充了服从对称性的有效场论概念,这些概念是对非扰动的小尺度效应的有效粗粒度描述。产生的有效描述具有与真实物理学类似的结构,但是具有必须满足的自由系数,并说明了非扰动的小尺度效应。
1. D. J. Eisenstein et al., SDSS-III: Massive spectroscopic surveys of the distant universe, the milky way, and extra-solar planetary systems. Astron. J. 142, 72 (2011).2. Z. Ivezic et al., LSST: From science drivers to reference design and anticipated data products. Bull. Am. Astron. Soc. 41, 366 (2009).3. D. Spergel et al., Wide-field infrarred survey telescope-astrophysics focused tele- scope assets WFIRST-AFTA 2015 report. arXiv Preprint. https://arxiv.org/abs/ 1503.03757 (Accessed 6 April 2021).4. A. Aghamousa et al., The DESI experiment part I: Science, targeting, and sur- vey design. arXiv Preprint. https://arxiv.org/abs/1611.00036 (Accessed 6 April 2021).5. L. Amendola et al., Cosmology and fundamental physics with the Euclid satellite. Living Rev. Relat. 21, 2 (2018).6. V. Springel et al., Simulations of the formation, evolution and clustering of galaxies and quasars. Nature 435, 629–636 (2005).7. J. DeRose et al., The Aemulus project. I. Numerical simulations for precision cosmology. Astrophys. J. 875, 69 (2019).8. A. A. Berlind, D. H. Weinberg, The halo occupation distribution: Toward an empiri- cal determination of the relation between galaxies and mass. Astrophys. J. 575, 587 (2002).9. P. Behroozi, R. H. Wechsler, A. P. Hearin, C. Conroy, Universemachine: The correlation between galaxy growth and dark matter halo assembly from z = 0-10. Mon. Not. R. Astron. Soc. 488, 3143–3194 (2019).10. I. Goodfellow et al., “Generative adversarial nets” in Advances in Neural Infor- mation Processing Systems, Z. Ghahramani, M. Welling, C. Cortes, N. Lawrence, K. Q. Weinberger, Eds. (MIT Press, 2014), pp. 2672–2680.11.D. P. Kingma, M. Welling, Auto-encoding variational Bayes. arXiv Preprint. https://arxiv.org/abs/1312.6114 (Accessed 4 April 2021).12. D. J. Rezende, S. Mohamed, D. Wierstra, Stochastic backpropagation and approx- imate inference in deep generative models. arXiv Preprint. https://arxiv. org/abs/1401.4082 (Accessed 6 April 2021).13. C. Modi, Y. Feng, U. Seljak, Cosmological reconstruction from galaxy light: Neural network based light-matter connection. J. Cosmol. Astropart. Phys. 2018, 028 (2018).14. P. Berger, G. Stein, A volumetric deep convolutional neural network for simulation of mock dark matter halo catalogues. Mon. Not. R. Astron. Soc. 482, 2861–2871 (2019).15. D. K. Ramanah, T. Charnock, G. Lavaux, Painting halos from cosmic density fields of dark matter with physically motivated neural networks. Phys. Rev. D 100, 043515 (2019).16. M. Bernardini, L. Mayer, D. Reed, R. Feldmann, Predicting dark matter halo formation in n-body simulations with deep regression networks. Mon. Not. R. Astron. Soc. 496, 5116–5125 (2020).17. X. Zhang et al., From dark matter to galaxies with convolutional networks. arXiv Preprint. https://arxiv.org/abs/1902.05965 (Accessed 6 April 2021).18. T.Tro ̈ster,C.Ferguson,J.Harnois-De ́raps,I.G.McCarthy,Paintingwithbaryons:Aug- menting n-body simulations with gas using deep generative models. Mon. Not. R. Astron. Soc. Lett. 487, L24–L29 (2019).19. F. List, I. Bhat, G. F. Lewis, A black box for dark sector physics: Predicting dark matter annihilation feedback with conditional GANs. Mon. Not. R. Astron. Soc. 490, 3134– 3143 (2019).20.E. Giusarma et al., Learning neutrino effects in cosmology with convolutional neural networks. arXiv Preprint. https://arxiv.org/abs/1910.04255 (Accessed 6 April 2021).21. D. K. Ramanah, T. Charnock, F. Villaescusa-Navarro, B. D. Wandelt, Super-resolution emulator of cosmological simulations using deep physical models. Mon. Not. R. Astron. Soc. 495, 4227–4236 (2020).22. Y. Li et al., AI-assisted super-resolution cosmological simulations. arXiv Preprint. https://arxiv.org/abs/2010.06608 (Accessed 6 April 2021).23. S. Tassev, M. Zaldarriaga, D. J. Eisenstein, Solving large scale structure in ten easy steps with Cola. J. Cosmol. Astropart. Phys. 2013, 036 (2013).24. Y. Feng, M.-Y. Chu, U. Seljak, P. McDonald, FastPM: A new scheme for fast simulations of dark matter and haloes. Mon. Not. R. Astron. Soc. 463, 2273–2286 (2016).25. S. He et al., Learning to predict the cosmological structure formation. Proc. Natl. Acad. Sci. U.S.A. 116, 13825–13832 (2019).26. A. Pillepich et al., First results from the IllustrisTNG simulations: The stellar mass content of groups and clusters of galaxies. Mon. Not. R. Astron. Soc. 475, 648–675 (2018).27. J. P. Naiman et al., First results from the IllustrisTNG simulations: A tale of two elements–chemical evolution of magnesium and europium. Mon. Not. R. Astron. Soc. 477, 1206–1224 (2018).28. F. Marinacci et al., First results from the IllustrisTNG simulations: Radio haloes and magnetic fields. Mon. Not. R. Astron. Soc. 480, 5113–5139 (2018).29. V. Springel et al., First results from the IllustrisTNG simulations: Matter and galaxy clustering. Mon. Not. R. Astron. Soc. 475, 676–698 (2018).30. D. Nelson et al., First results from the IllustrisTNG simulations: The galaxy colour bimodality. Mon. Not. R. Astron. Soc. 475, 624–647 (2018).31. J. J. M. Carrasco, M. P. Hertzberg, L. Senatore, The effective field the- ory of cosmological large scale structures. J. High Energy Phys. 2012, 82 (2012).32. B. Dai, Y. Feng, U. Seljak, A gradient based method for modeling baryons and matter in halos of fast simulations. J. Cosmol. Astropart. Phys. 2018, 009 (2018).33. B. Dai, Y. Feng, U. Seljak, S. Singh, High mass and halo resolution from fast low resolution simulations. J. Cosmol. Astropart. Phys. 2020, 002 (2020).34. D. H. Rudd, A. R. Zentner, A. V. Kravtsov, Effects of baryons and dissipation on the matter power spectrum. Astrophys. J. 672, 19 (2008).35. M. P. van Daalen, J. Schaye, C. M. Booth, C. D. Vecchia, The effects of galaxy formation on the matter power spectrum: A challenge for precision cosmology. Mon. Not. R. Astron. Soc. 415, 3649–3665 (2011).36. H.-J. Huang, T. Eifler, R. Mandelbaum, D. Scott, Modelling baryonic physics in future weak lensing surveys. Mon. Not. R. Astron. Soc. 488, 1652–1678 (2019).37.U.Seljak,G.Aslanyan,F.Yu,C.Modi, Towards optimal extraction of cosmological information from nonlinear data. J. Cosmol. Astropart. Phys. 2017, 009 (2017).38. R. H. Byrd, P. Lu, J. Nocedal, C. Zhu, A limited memory algorithm for bound constrained optimization. SIAM J. Sci. Comput. 16, 1190–1208 (1995).39. D. Nelson et al., The IllustrisTNG simulations: Public data release. Comput. Astrophys. Cosmol. 6, 1–29 (2019).