dave1368/cluster-06-bubble-dynamics-pinn

Cluster 6 Bubble Dynamics PINN

A Physics-Informed Neural Network predicting bubble radius R(t) during spherical cavity collapse, at any initial radius and driving acoustic pressure. Trained as part of the 9-cluster Scientific AI Cluster Orchestration Framework, which pairs this network with an exact symbolic ("Symetria") Rayleigh collapse-time solver and two physics-grounded safety audits under LangGraph supervision.

Architecture

The Rayleigh collapse is self-similar — x(τ) = R/R0 as a function of τ = t/Tc(R0,ΔP) is a universal curve, independent of the specific R0/ΔP sampled (the same structural fact as Cluster 4's Blasius similarity variable). The network is given τ (computed internally using the same exact Tc formula Symetria's own solver uses) plus log-scaled R0/ΔP (both span ~500x on their Gradio sliders), and predicts the universal x value directly.

Quickstart

import torch
from huggingface_hub import hf_hub_download
from modeling import BubbleDynamicsPINN

ckpt_path = hf_hub_download("dave1368/cluster-06-bubble-dynamics-pinn", "bubble_dynamics_pinn.pt")
# weights_only=False: the checkpoint is a dict with metadata (model_state_dict
# plus training info), not a bare tensor, so torch's default-safe loader can't
# be used as-is. Only do this for checkpoints you trust the source of.
checkpoint = torch.load(ckpt_path, map_location="cpu", weights_only=False)

model = BubbleDynamicsPINN()
model.load_state_dict(checkpoint["model_state_dict"])  # checkpoint also carries training-time loss history, see training_metrics.json
model.eval()

# inputs: (time_seconds, initial_radius_meters, driving_pressure_pascals)
inputs = torch.tensor([[1.0e-6, 50e-6, 1e6]])
radius_m = model(inputs)
print(radius_m)  # tensor([[R]])

Training data

Exact self-similar Rayleigh cavity-collapse solution — no synthetic correlation needed. The universal ODE dx/dτ = −K·√(x⁻³−1) (K=RAYLEIGH_CONST·√(2/3)≈0.74683) is integrated once via RK4 from x=1 (τ=0) down to a small floor (avoiding the model's genuine finite-time singularity at R→0), then reused for every (t, R0, ΔP) sample via interpolation:

• —60,000 training points, 10,000 validation points
• —Domain: R0 ∈ [1, 500] μm, ΔP ∈ [10, 5000] kPa, τ oversampled near the collapse-completion point
• —Final train loss: 2.02e-05 · Final val loss: 2.06e-05 (MSE, 3000 epochs)

Validated against classical sources (post-deployment finding)

Cross-checked against Rayleigh (1917), Plesset (1949), and Keller & Miksis (1980) — the papers cited in this cluster's Master Specification. Full data tables in the Space README.

A genuine, documented finding: the Mach-limit audit — meant to catch when collapse wall velocity exceeds water's sound speed — never fires anywhere in the practical input range, even at the most extreme slider corner (R0=1μm, ΔP=5000kPa, max predicted velocity 818 m/s vs. a 1481 m/s threshold). The underlying physics genuinely does go supersonic near true collapse for any input (the idealized Rayleigh model has a finite-time singularity at R→0), but a smooth neural network cannot represent that divergence — its fitted curve necessarily flattens out before reaching it. This is a function-class limitation already identified during training and not chased with more epochs.

Limitations

• —The Rayleigh model has a genuine finite-time singularity as R→0 — real physics (it's why bubble collapse produces shock waves and sonoluminescence), not a numerical artifact, and this network cannot represent it.
• —Gas-free (vapor-cavity) Rayleigh collapse only — no gas cushion, viscosity, or surface tension effects past the collapse point.
• —The Mach-limit audit does not currently discriminate in practice for any input in this app's slider range (see finding above).

Related

• —Space (interactive demo)