File size: 18,841 Bytes
fc7d689 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 | """Variational Autoencoder for diverse structural form-finding.
Implementation of a VAE coupled with a differentiable Force Density
Method (FDM) decoder for generating diverse equilibrium solutions.
The key insight: the FDM decoder is differentiable (via JAX implicit
differentiation), so the reparameterization trick (Kingma & Welling, 2014)
enables end-to-end training. The decoder is NOT modified -- it remains
the exact physics solver, guaranteeing equilibrium for every sample.
Mathematical formulation:
Encoder: mu, log_sigma = E_phi(X_hat)
Sampling: z = mu + exp(log_sigma) * epsilon, epsilon ~ N(0, I)
Mapping: q = (softplus(z) + tau) * s
Decoder: X(q) = K(q)^{-1} P (FDM equilibrium, unchanged)
Loss: L = L_shape(X, X_hat) + beta * KL(q(z|X_hat) || p(z))
Where:
- KL divergence: Eq. 7 of Kingma & Welling (2014), arXiv:1312.6114
- Beta annealing: Cyclical schedule per Fu et al. (2019), NAACL
- Physics guarantee: FDM enforces R(X;q) = 0 by construction
Motivation (from Pastrana et al., ICLR 2025, Section 6.1):
"the choice of bar stiffnesses for a given structure is not unique
and it is potentially appealing to present to the designer a diversity
of possible solutions by reformulating our model in a variational
setting (Kingma and Welling, 2014)"
References
----------
[1] Kingma, D.P. & Welling, M. (2014). Auto-Encoding Variational Bayes.
ICLR 2014. arXiv:1312.6114
[2] Fu, H. et al. (2019). Cyclical Annealing Schedule: A Simple Approach
to Mitigating KL Vanishing. NAACL 2019.
[3] Higgins, I. et al. (2017). beta-VAE: Learning Basic Visual Concepts
with a Constrained Variational Framework. ICLR 2017.
[4] Pastrana, R. et al. (2025). Real-Time Design of Architectural Structures
with Differentiable Mechanics and Neural Networks. ICLR 2025.
"""
from __future__ import annotations
import equinox as eqx
import jax
import jax.numpy as jnp
import jax.random as jrn
from jaxtyping import Array, Float, PRNGKeyArray
# =============================================================================
# Variational Encoder
# =============================================================================
class VariationalMLPEncoder(eqx.Module):
"""Variational MLP encoder for structural form-finding.
Maps target shapes to a diagonal Gaussian distribution in latent space,
then samples and maps to valid force densities via softplus + sign.
Architecture:
x -> [backbone MLP] -> h -> [mu_head] -> mu
-> [log_sigma_head] -> log_sigma
z = mu + exp(log_sigma) * epsilon (reparameterization trick [1])
q = (softplus(z) + q_shift) * edges_signs
The backbone shares features between mu and log_sigma heads,
improving parameter efficiency (standard VAE practice [1]).
Parameters
----------
backbone : eqx.nn.MLP
Shared feature extractor (depth-1 hidden layers).
mu_head : eqx.nn.Linear
Maps features to mean vector (unconstrained).
log_sigma_head : eqx.nn.Linear
Maps features to log-std vector. Bias initialized to -2.0
to start with small variance (sigma ~ 0.135), preventing
initial noise from destabilizing FDM [3, Section 4.1].
edges_signs : Array
+1 for tension, -1 for compression per edge.
q_shift : float
Minimum force density magnitude (tau in paper [4]).
"""
backbone: eqx.nn.MLP
mu_head: eqx.nn.Linear
log_sigma_head: eqx.nn.Linear
edges_signs: Array
q_shift: Float
slice_out: bool
slice_indices: Array
def __init__(
self,
edges_signs,
q_shift=0.0,
slice_out=False,
slice_indices=None,
in_size=300,
out_size=180,
width_size=256,
depth=3,
activation=jax.nn.elu,
*,
key,
):
k1, k2, k3 = jrn.split(key, 3)
# Shared backbone: depth-1 hidden layers
# Output is width_size features fed to both heads
self.backbone = eqx.nn.MLP(
in_size=in_size,
out_size=width_size,
width_size=width_size,
depth=max(depth - 1, 1),
activation=activation,
key=k1,
)
# Mean head: no activation (unconstrained)
self.mu_head = eqx.nn.Linear(width_size, out_size, key=k2)
# Log-sigma head: bias initialized to -2.0 for small initial variance
# This is critical to prevent the FDM decoder from receiving
# highly noisy q values at the start of training [3]
self.log_sigma_head = eqx.nn.Linear(width_size, out_size, key=k3)
# Override bias initialization
new_bias = jnp.full((out_size,), -2.0)
self.log_sigma_head = eqx.tree_at(
lambda l: l.bias, self.log_sigma_head, new_bias
)
self.edges_signs = edges_signs
self.q_shift = q_shift
self.slice_out = slice_out if slice_out else False
self.slice_indices = slice_indices
def __call__(
self,
x: Float[Array, "N3"],
*,
key: PRNGKeyArray | None = None,
) -> tuple[Float[Array, "E"], Float[Array, "E"], Float[Array, "E"]]:
"""Encode target shape to force density distribution and sample.
Parameters
----------
x : Array
Flat target shape (N*3,).
key : PRNGKey or None
Random key for sampling. If None, uses deterministic MAP
estimate z = mu (no sampling).
Returns
-------
q : Array (E,)
Force densities (physically valid: correct signs and shift).
mu : Array (E,)
Mean of approximate posterior q(z|x).
log_sigma : Array (E,)
Log standard deviation of approximate posterior.
"""
# Optional input slicing (same as Encoder, models.py:273-276)
if self.slice_out:
x = jnp.reshape(x, (-1, 3))
x = x[self.slice_indices, :]
x = jnp.ravel(x)
# Shared feature extraction
h = self.backbone(x)
# Distribution parameters
mu = self.mu_head(h)
log_sigma = self.log_sigma_head(h)
# Numerical stability: clamp log_sigma to prevent
# sigma explosion (>7.4) or exact zero (no KL gradient)
log_sigma = jnp.clip(log_sigma, -10.0, 2.0)
# Reparameterization trick (Kingma & Welling 2014, Eq. 4):
# z = mu + sigma * epsilon, where epsilon ~ N(0, I)
# This makes the sampling differentiable w.r.t. mu and sigma
if key is not None:
epsilon = jrn.normal(key, shape=mu.shape)
z = mu + jnp.exp(log_sigma) * epsilon
else:
# Deterministic mode: MAP estimate (no sampling)
z = mu
# Map to valid force densities
# softplus ensures positivity, then shift and sign are applied
# (same convention as MLPEncoder, models.py:332)
q = (jax.nn.softplus(z) + self.q_shift) * self.edges_signs
return q, mu, log_sigma
# =============================================================================
# Variational Autoencoder
# =============================================================================
class VariationalAutoEncoder(eqx.Module):
"""Variational autoencoder with differentiable FDM decoder.
Couples a variational encoder with the physics-based FDM decoder.
The decoder is NOT modified -- equilibrium is guaranteed for every
sample from the approximate posterior.
This enables generation of diverse equilibrium solutions from a
single target shape, addressing the non-uniqueness of force density
solutions noted in Pastrana et al. (2025), Section 6.1.
Parameters
----------
encoder : VariationalMLPEncoder
Variational encoder producing (q, mu, log_sigma).
decoder : FDDecoder
Physics-based decoder (unchanged from deterministic model).
"""
encoder: VariationalMLPEncoder
decoder: eqx.Module
def __init__(self, encoder, decoder):
self.encoder = encoder
self.decoder = decoder
def __call__(
self,
x: Float[Array, "N3"],
structure,
aux_data: bool = False,
*args,
key: PRNGKeyArray | None = None,
**kwargs,
):
"""Forward pass: encode, sample, decode.
Parameters
----------
x : Array
Flat target shape.
structure : EquilibriumStructure
Mesh structure.
aux_data : bool
If True, return auxiliary data for loss computation.
key : PRNGKey or None
Random key for reparameterization sampling.
Returns
-------
x_hat : Array
Predicted equilibrium shape.
vae_data : tuple (only when aux_data=True)
((q, xyz_fixed, loads), mu, log_sigma)
"""
from neural_fdm.gnn import GNNEncoder, VariationalGNNEncoder
if isinstance(self.encoder, (GNNEncoder, VariationalGNNEncoder)):
q, mu, log_sigma = self.encoder(x, structure=structure, key=key)
else:
q, mu, log_sigma = self.encoder(x, key=key)
x_hat = self.decoder(q, x, structure, aux_data)
if aux_data:
x_hat, params = x_hat # params = (q, xyz_fixed, loads)
return x_hat, (params, mu, log_sigma)
return x_hat
def encode(self, x, *, key=None, structure=None):
"""Encode target to distribution parameters."""
from neural_fdm.gnn import GNNEncoder, VariationalGNNEncoder
if isinstance(self.encoder, (GNNEncoder, VariationalGNNEncoder)):
return self.encoder(x, structure=structure, key=key)
return self.encoder(x, key=key)
def decode(self, q, *args, **kwargs):
"""Decode force densities to equilibrium shape."""
return self.decoder(q, *args, **kwargs)
def sample(
self,
x: Float[Array, "N3"],
structure,
key: PRNGKeyArray,
num_samples: int = 10,
) -> tuple[Float[Array, "S N3"], Float[Array, "S E"]]:
"""Generate diverse equilibrium shapes from a single target.
Samples multiple z values from q(z|x) and decodes each through
the FDM solver. Every sample is guaranteed to be in equilibrium.
Parameters
----------
x : Array
Single target shape (flat).
structure : EquilibriumStructure
Mesh structure.
key : PRNGKey
Random key.
num_samples : int
Number of diverse solutions to generate.
Returns
-------
x_hats : Array (num_samples, N*3)
Diverse equilibrium shapes.
qs : Array (num_samples, E)
Corresponding force densities.
"""
keys = jrn.split(key, num_samples)
def _sample_one(k):
from neural_fdm.gnn import GNNEncoder, VariationalGNNEncoder
if isinstance(self.encoder, (GNNEncoder, VariationalGNNEncoder)):
q, _, _ = self.encoder(x, structure=structure, key=k)
else:
q, _, _ = self.encoder(x, key=k)
x_hat = self.decoder(q, x, structure, False)
return x_hat, q
x_hats, qs = jax.vmap(_sample_one)(keys)
return x_hats, qs
def predict_states(self, x, structure):
"""Deterministic prediction for visualization.
Uses MAP estimate (key=None) for compatibility with the
existing visualization pipeline.
"""
x_hat, (params, mu, log_sigma) = self(x, structure, True, key=None)
from neural_fdm.models import build_states
return build_states(x_hat, params, structure)
# =============================================================================
# KL Divergence
# =============================================================================
def compute_kl_divergence(
mu: Float[Array, "... D"],
log_sigma: Float[Array, "... D"],
) -> Float[Array, ""]:
"""KL divergence between diagonal Gaussian and standard normal.
KL(q(z|x) || p(z)) where:
q(z|x) = N(mu, diag(sigma^2))
p(z) = N(0, I)
Formula (Kingma & Welling 2014, Appendix B, Eq. 7):
KL = -0.5 * sum_j (1 + log(sigma_j^2) - mu_j^2 - sigma_j^2)
= -0.5 * sum_j (1 + 2*log_sigma_j - mu_j^2 - exp(2*log_sigma_j))
Parameters
----------
mu : Array (..., D)
Mean of approximate posterior.
log_sigma : Array (..., D)
Log standard deviation of approximate posterior.
Returns
-------
kl : scalar
Mean KL divergence over batch.
References
----------
[1] Kingma & Welling (2014), arXiv:1312.6114, Eq. 7
"""
kl_per_dim = -0.5 * (
1.0 + 2.0 * log_sigma - jnp.square(mu) - jnp.exp(2.0 * log_sigma)
)
# Sum over latent dimensions (per-sample KL), then mean over batch.
# Consistent for both single sample (1D) and batch (2D):
# - 1D: sum reduces to scalar, mean of scalar = scalar
# - 2D: sum per row, mean over rows
kl_per_sample = jnp.sum(kl_per_dim, axis=-1)
return jnp.mean(kl_per_sample)
# =============================================================================
# Beta Annealing Schedule
# =============================================================================
def compute_beta_schedule(
step: int,
beta_max: float = 1.0,
cycle_length: int = 5000,
warmup_ratio: float = 0.5,
) -> float:
"""Cyclical beta annealing schedule.
Beta linearly increases from 0 to beta_max during the warmup
portion of each cycle, then stays at beta_max for the rest.
This prevents posterior collapse by allowing the encoder to
first learn a good reconstruction, then gradually enforce
the prior constraint.
Schedule per Fu et al. (2019):
beta(t) = beta_max * min(1, (t mod T) / (T * r))
where T = cycle_length, r = warmup_ratio.
Parameters
----------
step : int
Current training step.
beta_max : float
Maximum beta value. beta=1.0 gives standard VAE ELBO.
beta<1.0 allows reconstruction to dominate (underfitting prior).
beta>1.0 gives beta-VAE (Higgins et al. 2017) for stronger
disentanglement.
cycle_length : int
Number of steps per annealing cycle.
warmup_ratio : float
Fraction of cycle for linear warmup (0 to 1).
Returns
-------
beta : float
Current beta value in [0, beta_max].
References
----------
[2] Fu et al. (2019). Cyclical Annealing Schedule. NAACL 2019.
[3] Higgins et al. (2017). beta-VAE. ICLR 2017.
"""
position = step % cycle_length
warmup_length = cycle_length * warmup_ratio # float division, no rounding
beta = beta_max * min(1.0, position / max(warmup_length, 1e-10))
return beta
# =============================================================================
# Solution Multiplicity Metrics
# =============================================================================
def compute_diversity_metrics(
x_hats: Float[Array, "S N3"],
qs: Float[Array, "S E"],
) -> dict:
"""Quantify solution multiplicity from VAE samples.
Given S samples of equilibrium shapes and force densities from the
same target, computes metrics characterizing the diversity of solutions.
Quantifies the force density solution multiplicity documented
qualitatively by Veenendaal & Block (2012) and Adriaenssens et al. (2014).
Parameters
----------
x_hats : Array (S, N*3)
S sampled equilibrium shapes from the same target.
qs : Array (S, E)
Corresponding force density vectors.
Returns
-------
metrics : dict
- "n_samples": number of samples
- "shape_pairwise_L1_mean": mean pairwise L1 distance between shapes
- "shape_pairwise_L1_std": std of pairwise L1 distances
- "q_pairwise_L1_mean": mean pairwise L1 distance between q vectors
- "q_std_per_edge": Array (E,) std of q across samples per edge
- "q_std_mean": mean of per-edge std (scalar summary)
- "shape_std_per_node": Array (N,) std of position across samples per node
References
----------
Veenendaal & Block (2012). "An overview and comparison of structural
form finding methods." IJSS, 49(26):3741-3753.
Adriaenssens et al. (2014). Shell Structures for Architecture. Routledge.
"""
S = x_hats.shape[0]
# Pairwise L1 distances (shapes)
shape_dists = []
q_dists = []
for i in range(S):
for j in range(i + 1, S):
shape_dists.append(float(jnp.sum(jnp.abs(x_hats[i] - x_hats[j]))))
q_dists.append(float(jnp.sum(jnp.abs(qs[i] - qs[j]))))
import numpy as np
shape_dists = np.array(shape_dists)
q_dists = np.array(q_dists)
# Per-edge q standard deviation (where is there freedom?)
q_std = np.std(np.array(qs), axis=0)
# Per-node shape standard deviation
x_reshaped = np.array(x_hats).reshape(S, -1, 3)
shape_std = np.std(np.linalg.norm(x_reshaped, axis=-1), axis=0)
return {
"n_samples": S,
"shape_pairwise_L1_mean": float(np.mean(shape_dists)) if len(shape_dists) > 0 else 0.0,
"shape_pairwise_L1_std": float(np.std(shape_dists)) if len(shape_dists) > 0 else 0.0,
"q_pairwise_L1_mean": float(np.mean(q_dists)) if len(q_dists) > 0 else 0.0,
"q_std_per_edge": q_std,
"q_std_mean": float(np.mean(q_std)),
"shape_std_per_node": shape_std,
}
def compute_variance_per_edge(
model,
x_target: Float[Array, "N3"],
structure,
key,
n_samples: int = 50,
) -> Float[Array, "E"]:
"""Map solution freedom to individual structural members.
Samples n_samples force density vectors from the VAE posterior
and computes the variance of q per edge. High variance indicates
the structure has design freedom at that member; low variance
indicates the member is strongly constrained by the target geometry.
This enables visualization of "structural design freedom" per member,
a capability enabled by the variational formulation.
Parameters
----------
model : VariationalAutoEncoder
Trained VAE model.
x_target : Array (N*3,)
Target shape.
structure : EquilibriumStructure
Mesh structure.
key : PRNGKey
Random key for sampling.
n_samples : int
Number of samples (more = better estimate).
Returns
-------
q_variance : Array (E,)
Variance of force density per edge across samples.
"""
_, qs = model.sample(x_target, structure, key, num_samples=n_samples)
return jnp.var(qs, axis=0)
|