AM227 物理中的數值方法

AM227 物理中的數值方法

來自專欄 Insomnia in Boston1 人贊了文章

AM227 Computational Methods in the Physical Sciences,這學期選的一門課,主要介紹FDM、FVM、Lattice Boltzmann method,以及Machine Learning與求解PDE的結合應用。公開的課程介紹在

Computational Physics?

projects.iq.harvard.edu圖標

更詳細的版本是:

In this Course, we shall familiarize with the main computational methods which permit to simulate and analyze the behavior of a wide range of problems involving fluids, solids, soft matter, electromagnetic and quantum systems, as well as the dynamics of (some) biological and social systems.

The course consists of three main parts,

Part I : Fields on Grids

Part II : Mesoscale Particle Methods

Part III: Statistical Data Analysis and Learning

In Part I, we shall discuss the fundamentals of grid discretization and present concrete applications to a broad variety of problems from classical and quantum physics, such as Advection-Diffusion Reaction transport, Navier-Stokes fluid-dynamics, nonlinear classical and quantum wave propagation. Both regular and complex geometrical grids will be discussed through Finite Differences, Volumes and Elements, respectively.

In Part II we shall discuss mesoscale technique based on the two basic mesoscale descriptions: probability distribution functions, as governed by Boltzmann and Fokker-Planck kinetic equations, and stochastic particle dynamics (Langevin equations). The lattice Boltzmann method will be discussed in great detail, with applications to fluids and soft matter problems. Mesoscale particle methods, such as Dissipative Particle Dynamics will also be illustrated in detail.

Finally, in Part III, we shall present data analysis & learning tools of particular relevance to complex systems with non-gaussian statistics, such as turbulence, fractional transport and extreme events in general. An introduction to Physics-Aware Machine Learning will also be presented.

老師叫Sauro Succi

Sauro Succi - Wikipedia?

en.wikipedia.org

,義大利人,主要做流體的數值計算,2001年發表了一本書《The Lattice Boltzmann Equation: For Fluid Dynamics and Beyond》,現在應該已經退休了,2014年開始來Harvard IACS當visiting professor。

本篇先寫點開學初的課程介紹部分。


Why Computational Physics?

  • Experiment和Theory有固有的限制(老生常談了)
  • Computational Physics的兩大基石:硬體和軟體。
  1. 硬體滿足Moores Law的增長,1980s→Mflops,1990s→Gflops,2000s→Tflops,2010s→Pflops,那麼2020s→Eflops?
  2. 軟體層面主要取決於演算法的創新: e_p approx frac{1}{N^p} ,N是網格數,p是階數,e是誤差。比如,1000個節點的二階網格,取代1000*1000節點的一階網格,相當於節約了15個Moore年。
  • 十大演算法:
  1. Monte Carlo:高維空間
  2. Molecular Dynamics:凝聚態物質,生物學
  3. CFD:流體
  4. FFT:信號
  5. Fortran編譯器:所有方程
  6. QR分解:線性代數
  7. Lanczos/Krylov iteration:迭代
  8. Fast Multipole:長程相互作用
  9. Simplex method:線性規劃
  10. Quicksort:搜索
  • 當下求解PDE的主要困難:非線性,非局域性,高維,幾何複雜性
  • 統計力學角度看描述物質的方程:
  1. Quantum:薛定諤方程
  2. Micro:分子層面的經典力學方程
  3. Meso:玻爾茲曼方程
  4. Macro:連續介質的NS方程
  • 連續介質場和PDE
    • Navier-Stokes queations以及Turbulence
      • Mass conservation: partial_t
ho+Deltacdot(
ho vec{u})=0
      • Momentum conservation: partial_t(
hovec{u})+Deltacdot(
ho vec{u} vec{u} + overleftrightarrow{P})=vec{f}
    • Solid Mechanics
    • Electromagnetics
    • Quantum Physics
    • Computational Chem/Bio
    • Quantum Field Theory
  • 複雜幾何上的場:FVM、FEM
  • 從場到粒子:PDE到ODE
    • Molecular Dynamics來自於牛頓力學
    • 玻爾茲曼動力學
      • 比如6維相空間來來描述流體: partial_tf+vec{v}cdot Delta_rf+(vec{F}/m)cdot Delta_vf=C(f,f)
      • 非常難求解,高維,非線性
      • 可以描述很多現象:中子運輸、γ射線傳輸、太空梭回收、電子流,以及人類的交通流
      • 晶格玻爾茲曼方法《The Lattice Boltzmann Equation》by Sauro Succi
      • Fokker-Planck方程,最重要的統計力學方程質疑,用於Chemical-physics、量子力學,以及社會科學,因為處理了隨機過程。 partial_tP+Deltacdot vec{J}[P] = 0\ P(q,t):Probability density function

Big data for computational physics

  • All models are wrong, but some are useful (G.Box) → All models are wrong and useless :)哈哈。我們需要認識到Big data的意義。
  • Big Data需要注意的4點:(SS and P.Coveney, Phil. Trans. Roy. Soc.)
    • 1)複雜系統是高度相關的:不滿足高斯統計,因此誤差下降很慢,甚至可能隨著數據增加而上升
    • 2)相關≠因果
    • 3)太多數據和沒有數據一樣糟糕
    • 4)數據本身的誤差和模糊
  • 複雜系統例如CFD的turbulence,氣候,地震,stock price的波動不符合高斯過程
  • 小概率事件的危害可能很大,Impact=Probability*Intensity:比如黑天鵝事件
  • 相關和因果無關的例子:

15 Insane Things That Correlate With Each Other?

tylervigen.com圖標

  • 數據量和信息量的關係:data starving→data driven→data saturation→data buried
    • seeing everything is like seeing nothing。。。
  • 數據誤差的蝴蝶效應:誤差的增長指數增加,data chaos
  • 物理與機器學習!!!
    • 物理輔助的機器學習(physics assisted machine learning):用模擬環境替代實際環境來做預測和有監督學習。比如ML for fluid turbulence。
    • 機器學習輔助的模擬(ML-assisted simulations):
      • 比如 2015 PRL E.D. Cubuk :Identifying Structural Flow Defects in Disordered Solids Using Machine-Learning Methods
    • 機器學習做精確製藥(?)
      • Simulating Soft-Sphere Margination in Arterioles and Venules

推薦閱讀:

LLVM每日談之二十九 面向機器學習的編譯器——Glow
機器學習筆記四--邏輯回歸
機器學習入門 - 邏輯回歸
Patchouli的機器學習系列教程六:非監督學習——『器』
初學機器學習:直觀解讀KL散度的數學概念

TAG:物理學 | 機器學習 | 模擬 |