Machine-learning Construction of a Model and Refined Regularity Criterion on Fluid Equations
概要
大気や水の流れのような多くの流体運動の挙動は偏微分方程式である Navier–Stokes 方程式によって記述されることが知られている. この微分方程式は, 非局所性などからくる高い計算コストや, 強い非線形性などからくる数学的な難しさなど様々な難しさを持ち合わせている.
本論文では, 以下の 2 つの課題に取り組む.
· 流体マクロ変数の機械学習を用いた数理モデルの構成
· 一般化した Navier–Stokes 方程式の滑らかな解の延長問題
1 機械学習によるモデリング
3 次元 Navier–Stokes 方程式は以下の連立方程式である:
∂tv + (v · ∇)v = −∇π + ν∆v in R 3 × (0,∞)
div v = 0 in R 3 × (0, ∞)
v(x, 0) = v0(x) in R 3 .
ただし, v = v(x, t) = (v1(x, t), v2(x, t), v3(x, t)) を速度ベクトル, π = π(x, t) を圧力, ν > 0 を動粘性係数と する. また, v0(x) は div v0 = 0 を満たす初期速度ベクトルする. J. Leray (1934), E. Hopf (1951) らによっ て弱解の存在が示されてる. しかし, その弱解の正則性や一意性の有無はいまだに証明されていない. この, 大 きな初期値 v0 に対して滑らかな時間大域解が一意に存在するか否かという問題は 21 世紀に考えるべき問題 としてクレイ研究所が発表している 7 つのミレニアム問題うちの一つである.
この Navier–Stokes 方程式の時間大域解の存在を示す上で, 速度のエネルギー ∥v∥L2 などのマクロ変数の挙 動を知ることは重要である. しかし, 速度などのミクロ変数の方程式である Navier–Stokes 方程式からマクロ 変数で閉じた時間発展方程式を解析的に導出することが困難なことが知られている. 例えば, n 次のモーメン ト変数の支配方程式を書き下すためには n + 1 次モーメント変数が必要になり, 高次モーメント量を無視する ことができないという閉包問題がある. これは Navier–Stokes 方程式の非線形効果や非局所性などに由来する ものである. そのため, マクロ変数のみのデータを得る際も Navier–Stokes 方程式を解く必要がある. 流体の モデリング手法は開発されてきているが, その多くは非線形効果を人工的にある程度無視することで得ている.
そこで, 流体変数のモデリングに機械学習を使って取り組んだ. また, マクロ変数としてエネルギー変数や乱 流の複雑さを表すレイノルズ数についてモデルを構成した.
1.1 機械学習手法
本研究で主に扱う “学習” とは, ある時刻の M 次元変数のデータ u(t) ∈ RM からその次の時刻の変数の データ s(t + ∆t) ∈ RM を結ぶ対応関係 “F” (F : RM → RM) を見つけることである;
F(u(t)) ≈ s(t + ∆t) for any t ∈ [0, T].
もし “よい” 対応関係 F が見つかった場合, 新しい入力データ u(T + 1) を使って得られたデータ
s(T + 1 + ∆t) = F(u(T + 1)),
は s(T + 1 + ∆t) のよい予測になっているはずである.
しかし, 入力次元だけでは十分ではない場合がある. 特に本研究で取り扱うマクロ変数は, 背後にあるミクロ 変数の力学系の次元を低次元空間に射影したものである. この場合, 単にマクロ変数を入力変数として次の時 刻のマクロ変数との対応関係を得ることは困難であることは想像に固くないであろう. 例えば, ジェットコー スターのコースの影のみを見て軌跡を予想できるかという問題を考えると想像しやすいだろう. そこで, 入 力したものを適当に分解し高次元空間 R N (N >> M) に射影 F1(F1 : RM → R N ) して, その後適当な関数 F2(F2 : R N → RM) を使って出力を得るという 2 段階にし, F2(F1(u(t))) ≈ s(t + ∆t) を満たすように関数 F1, F2 を探すと良い.
さらに, 時系列データの学習において困難な点あり, それは sin 関数を例に挙げると分かりやすい. 例えば, sin(x) = 0(= u(x)) としたとき, s(x + π/2) = sin(x + π/2) は {±1} のどちらか定めることができない. この 出力を一意に定めるためには sin(x − π/2) の正負などの過去のデータも加味して学習する必要がある. そこ で, 射影 F1 を過去のデータも加味できるように変更し, F2(F3(u(t), r(t))) ≈ s(t + ∆t) を満たすように関数 F3, F2 を探すと良い. ただし, F3 : RM × R N → R N . r(t) = F3(u(t − ∆t), r(t − ∆t)) ∈ R N とする.
しかし, F3 を変えると F3(u(t), r(t)) で生成される時系列データが大きく影響を受けるため, 良い関数 F2, F3 を見つけるためには非常に計算コストがかかる. このため流体のように超高次元力学系の学習には適さ ない. そこで, 関数 F3 は適当に与えて, F2 のいい関数の発見に力を注ぐことにした学習手法はリザーバーコ ンピューティング (以下, RC 学習) と呼ばれる. 本論文の学習では RC 学習を用いる.
1.2 流体マクロ変数のモデリング
RC 学習は, 2017 年ごろから注目されるようになり, メリーランド大学の Prof. Hunt らによるローレン ツ方程式などの低自由度のトイモデルを用いた研究などがある. しかし, 大自由度な力学系であり現実の物理 現象を記述する流体運動に対する RC 学習は成功してはいない. そこで, まずは流体変数について, RC 学習 を用いてモデリングをした.
本論文の第 1 章では, マクロ変数としてエネルギー関数の時系列データを学習しモデリングした. この結果 は, 高次元力学系であるため従来の手法では機械学習が困難であった乱れた流体運動の時間発展モデリングに ついて, RC 学習を使うことで初めて学習に成功したものである. ただし, エネルギー関数 E0(k, t) は
E0(k, t) := 1 2 ∑ κ∈Dk ∑ 3 ζ=1 F[vζ](κ, t) 2 ,
として定義する. また, F を R 3 上の 2 乗可積分関数 L 2 から L 2 へのフーリエ変換として, Dk := {κ ∈ Z 3 |k − 0.5 ≤ |κ| < k + 0.5} は領域, k ∈ N は波数, t > 0 は時間を表す. 更に E0(k, t) の規格化などをしたも のを E˜(k, t) とする. このもとで,
u(t) = (E˜(1, t), E˜(2, t), · · · , E˜(9, t))t,
s(t) = (E˜(1, t), E˜(2, t), · · · , E˜(9, t))t,
として RC 学習をしてモデリングをした. これはつまり, ある時刻 t のエネルギー関数のデータ u(t) を入力す ると, 次の時刻 t + ∆t のエネルギー関数のデータ s(t + ∆t) を出力するモデルを構成することに相当する. こ の構成したモデルを用いて, 入力データ u(t + ∆t) として, 前の時刻で予測した出力データ s(t + ∆t) を代用す ることで, 出力データ s はエネルギー関数の時系列データを得られる. また, エネルギー関数の時間発展を予 測することに成功していることを確認した.
また, 学習可能な時系列データの数が流体運動の不安定次元より少ない数にとどまる場合にも, 時間遅れ座 標を導入することで十分な学習可能な時系列データがある場合と同レベルの予測精度がある数理モデルが構築 できることを確認した. このことは, 実用上は観測可能な時系列の数は少数に限定されることが通例であるが, そのような場合にも本手法を適用することが可能であることを意味する.
本論文の第 2 章では, 時間依存の流体運動の複雑さを表すテーラーマイクロスケールのレイノルズ数 Rˇ λ(t) について, 遅れ時間座標系を使うことで時系列データを学習しモデリングしている. レイノルズ数 Rˇ λ(t) は次 で定義する:
Rˇ λ(t) := √ 20E(t) 2 3νϵ(t) ,
ただし、
ϵ(t) = 2ν ∑ κ∈Z3 ∑ 3 ζ=1 |κ| 2 ( F[vζ](κ, t) )2 ,
は粘性の効果によるエネルギー散逸率, E(t) = ∑ k E0(k, t) は全エネルギーを意味する.
入力次元が十分でない場合は時間遅れ座標を導入することで精度よく数理モデルが構築可能であることを 1 章で確認した. 今回着目するレイノルズ数はスカラー値関数であるため, 本章でも時間遅れ座標を用いて学習 をし, その時間発展を予測することに成功していることを確認した. さらに効率良く学習をするための時間遅 れ座標の遅れ時間と座標系の次元の最適なとり方について自己相関係数の値に注目して考察した. この結果は, RC 学習においてよい遅れ時間座標系の取り方を詳細に調べた初めての結果である. また, 最も効率の良い時 間遅れ座標系を導入することにより速度場が穏やかな流れから複雑な流れに時間間欠的に遷移する様子も再現 可能であることも確認した.
2 解の延長問題
まず, 渦度 ω は速度 v, 微分作用素 rot を用いて次で定義する:
ω = rot v.
任意の T(< T ⋆ ) に対して Navier–Stokes 方程式の解の渦度が ω ∈ L 1 (0, T;L∞(R 3 )) を満たす時, Es(T ⋆ ) の 関数クラスで解がなめらかに延長できるということが知られている. ただし, 関数クラス Es(T) は, ソボレフ空間 Hs を使って次のように定義した:
Es(T) := C([0, T]; Hs ) ∩ C 1 ([0, T]; Hs−1 ) (s ≥ 3).
J. T. Beale, T. Kato, A. Majda (1984) らは Euler 方程式について解の延長問題を証明したが, 全く同じ証 明方法で, Navier–Stokes 方程式についても証明することができる. さらに, P. Constantin, C. Fefferman (1993) は渦度の方向ベクトルの連続性の指標
√ 1 − (ξ(x, t) · ξ(x + h, t))2 |h| β (=: ηβ(x, h, t)) (1)
を新たに導入した. 渦度の方向ベクトルの Lipschitz 連続性 (β = 1) を仮定すると Navier–Stokes 方程式の解 の正則性が得られることが示されている. これを, H. Beir˜ao da Veiga, L. Berselli (2002) らが β-H¨older 連 続性 (β = 1/2) あれば十分であると改良した.
一方で, Navier–Stokes 方程式の単純な拡張として次の連立方程式を考えることとした:
∂tv + (v · ∇)v = −∇π − ν(−∆)α/2 v in R 3 × (0,∞) (2)
div v = 0 in R 3 × (0, ∞) (3)
v(x, 0) = v0(x) in R 3 . (4)
として定義する.
この連立偏微分方程式 (??)∼(??) を (NS)α と書く. エネルギーの減衰則と呼ばれる Navier–Stokes 乱流の 重要な性質の一つについて, (NS)α でもこの減衰則を満たすようにできることが Z. Xiao らによって示されて いる (2009). また, (NS)α は J. L. Lions (1969) によて初めて研究され, α ≥ 5/2 での時間大域的な滑らかな 一意解の存在を示されている. 一方で, 主定理で考察する 0 < α < 2 のようにラプラシアンの冪が 1 未満の場 合では, 弱解の存在すら示されていない.
本論文の第 3 章では, 渦度 ω の属する関数クラスと渦度の方向ベクトルの連続性の指標 ηβ に基づいて, (NS)α の関数クラス Es(T) での解の延長問題を証明した.
1 b + 1 r < α + β 3 , α a + α q + 3 b + 3 r ≤ α + β,
として、
ω ∈ L q (0, T;L r ).
· ηβ ≤ g(x, t) ただし, g ∈ L a (0, T;L b ).
を満たすとき, 解が時刻 T ∗ > T まで関数クラス Es(T ∗ ) でなめらかに延長できる.
これは, Navier–Stokes 方程式から粘性の効果を弱めた方程式である (NS)α についても, Navier–Stokes 方 程式と同様にして渦度の方向ベクトルの β-H¨older 連続性が速度の滑らかさが得られることを示している.