图片详情
—
Skip to content

demo_lennard_jones.py ​

返回章节 · 配套代码说明

复制以下内容并保存为 demo_lennard_jones.py,使用方法见配套代码说明。

python
"""案例二:用神经网络学习 Lennard-Jones 双原子势能曲线。

这是「神经网络势函数(neural network potential)」的最小原型:
以原子间距 r 为输入,输出相互作用能 V。多原子神经网络势
(ANI、Behler-Parrinello、DeePMD……)还需要处理局域环境、对称性、受力和构型覆盖。

采用约化单位 ε = 1、σ = 1:

    V(r) = 4 [ (1/r)^12 - (1/r)^6 ]

极小值在 r_min = 2^(1/6) ≈ 1.12246,此时 V = -1。
"""

import math

from nn import MLP

EPS, SIG = 1.0, 1.0


def lj_energy(r):
    """Lennard-Jones 势能(约化单位)。"""
    x = SIG / r
    x6 = x ** 6
    return 4.0 * EPS * (x6 * x6 - x6)


def mean(v):
    return sum(v) / len(v)


def std(v):
    m = mean(v)
    return math.sqrt(sum((x - m) ** 2 for x in v) / len(v))


def rmse(a, b):
    return math.sqrt(sum((x - y) ** 2 for x, y in zip(a, b)) / len(a))


# ---- 训练数据:在 [0.95, 2.5] 上等间距取 48 个点 ----
def linspace(a, b, n):
    return [a + (b - a) * k / (n - 1) for k in range(n)]


r_train = linspace(0.95, 2.5, 48)
v_train = [lj_energy(r) for r in r_train]

# ---- 标准化(用训练集统计量) ----
r_mu, r_sd = mean(r_train), std(r_train)
v_mu, v_sd = mean(v_train), std(v_train)
X = [[(r - r_mu) / r_sd] for r in r_train]
Y = [[(v - v_mu) / v_sd] for v in v_train]


def std_r(r):
    return (r - r_mu) / r_sd


def unstd_v(z):
    return z * v_sd + v_mu


# ---- 基线:没有隐藏层 = 线性模型 ----
lin = MLP([1, 1], output_activation="linear", loss="mse", seed=0)
lin.train(X, Y, epochs=6000, lr=0.1, batch_size=len(X), momentum=0.9, seed=0)

# ---- 多层网络 ----
net = MLP([1, 16, 16, 1], hidden_activation="tanh",
          output_activation="linear", loss="mse", seed=0)
hist = net.train(X, Y, epochs=6000, lr=0.05, batch_size=len(X),
                 momentum=0.9, seed=0)

# ---- 评估:在更密的网格上比较 ----
grid = linspace(0.95, 2.5, 400)
true_v = [lj_energy(r) for r in grid]


def predict(model, r):
    return unstd_v(model.predict([std_r(r)])[0])


pred_lin = [predict(lin, r) for r in grid]
pred_net = [predict(net, r) for r in grid]

print("=" * 74)
print("用神经网络学习 Lennard-Jones 势能曲线(约化单位 ε=1, σ=1)")
print("=" * 74)
print(f"\n训练点:{len(r_train)} 个,r ∈ [0.95, 2.5]")
print(f"网络结构 [1, 16, 16, 1],参数量 {net.n_params()}")
print(f"训练损失:{hist[0]:.6f}  ->  {hist[-1]:.6f}")

print("\n在 400 点密集网格上的误差(约化能量单位)")
print(f"  线性模型  [1, 1]           RMSE = {rmse(true_v, pred_lin):.6f}")
print(f"  多层网络  [1, 16, 16, 1]   RMSE = {rmse(true_v, pred_net):.6f}")

# ---- 物理量的复现:极小值位置与深度 ----
def argmin(preds):
    k = min(range(len(preds)), key=lambda i: preds[i])
    return grid[k], preds[k]


r_true, v_true = argmin(true_v)
r_lin, v_lin = argmin(pred_lin)
r_net, v_net = argmin(pred_net)

print("\n势阱极小值的复现")
print(f"  {'':<12}{'r_min':>12}{'V_min':>12}")
print(f"  {'解析解':<12}{2 ** (1 / 6):>12.5f}{-1.0:>12.5f}")
print(f"  {'线性模型':<12}{r_lin:>12.5f}{v_lin:>12.5f}")
print(f"  {'多层网络':<12}{r_net:>12.5f}{v_net:>12.5f}")

# ---- 几个代表性间距的预测 ----
print("\n若干间距处的预测值")
print(f"  {'r':>6}{'真值':>12}{'线性':>12}{'多层网络':>12}")
for r in (1.00, 1.12, 1.30, 1.60, 2.00, 2.40):
    print(f"  {r:>6.2f}{lj_energy(r):>12.5f}{predict(lin, r):>12.5f}"
          f"{predict(net, r):>12.5f}")

# ---- 受力检验:对能量模型求导得到受力,并与解析受力比较 ----
def lj_force(r):
    """解析受力 F = -dV/dr(约化单位)。"""
    x6 = (SIG / r) ** 6
    return 24.0 * EPS / r * (2.0 * x6 * x6 - x6)


def net_force(model, r, h=1e-5):
    """用中心差分对网络预测的能量求导:F ≈ -[V(r+h) - V(r-h)] / 2h。"""
    return -(predict(model, r + h) - predict(model, r - h)) / (2.0 * h)


F_true = [lj_force(r) for r in grid]
F_net = [net_force(net, r) for r in grid]
inner = [k for k, r in enumerate(grid) if r >= 1.0]
print("\n由 MLP 能量求导得到的受力")
print(f"  受力 RMSE(r ∈ [0.95, 2.5])   = {rmse(F_true, F_net):.4f}")
print(f"  受力 RMSE(r ∈ [1.0, 2.5])    = "
      f"{rmse([F_true[k] for k in inner], [F_net[k] for k in inner]):.4f}")
print(f"  r = 0.95 处:解析 {lj_force(0.95):.2f},MLP {net_force(net, 0.95):.2f}")

print("\n训练区间之外的能量预测")
print(f"  {'r':>6}{'真值':>12}{'多层网络':>12}")
for r in (0.85, 0.90, 2.80, 3.20):
    print(f"  {r:>6.2f}{lj_energy(r):>12.4f}{predict(net, r):>12.4f}")

《原子智能》· 纸质出版预备版 · PolyAI Team 著