geometric-quantum-pinn

name

geometric-quantum-pinn

description

Geometric Quantum Physics-Informed Neural Network (GQPINN) methodology for solving PDEs with symmetry-aware quantum circuits. Combines geometric quantum machine learning with physics-informed neural networks. Use when solving PDEs with quantum circuits, incorporating symmetry/inductive biases into quantum models, or designing equivariant quantum ansatzes for scientific ML. Activation: geometric quantum, symmetry-aware PINN, quantum PDE solver, equivariant quantum circuit, GQPINN, quantum physics-informed neural network.

Geometric Quantum Physics-Informed Neural Network (GQPINN)

Overview

GQPINNs extend Quantum PINNs (QPINNs) by encoding the geometric structure of the underlying PDE directly into the quantum circuit ansatz. This produces symmetry-preserving predictions that respect the governing equation's invariances, achieving better accuracy with fewer trainable parameters.

Core Methodology

Step 1: Identify PDE Symmetries

Analyze the target PDE for symmetries:

• Finite group symmetries: Discrete transformations (reflections, permutations)
• Compact Lie group symmetries: Continuous transformations (rotations, translations)
• Scale invariances: Rescaling of independent/dependent variables

Step 2: Construct Equivariant Generator Sets

For each identified symmetry, build parametrized circuit generators:

# For a finite group G = {g_1, ..., g_k}:
# Build generators U(θ) that satisfy: U(θ) · ρ(g) = ρ(g) · U(θ)
# where ρ(g) is the group representation on the quantum state

# For Lie groups: use exponential map exp(i·θ·X) where X are Lie algebra generators

Step 3: Twirling-Based Symmetry Preservation

Apply group twirling to enforce symmetry:

U_sym(θ) = (1/|G|) Σ_{g∈G} ρ(g)† · U(θ) · ρ(g)

This ensures model predictions respect PDE symmetries when boundary/initial data are symmetry-compatible.

Step 4: Quantum Circuit Ansatz Design

Combine symmetry generators with trainable parameters:

• Equivariant layers: Apply symmetry-preserving transformations
• Physics loss terms: PDE residuals at collocation points
• Boundary conditions: Enforced via penalty or hard constraints

Step 5: Training Protocol

1. Sample collocation points from domain
2. Evaluate quantum circuit at each point → prediction u_θ(x,t)
3. Compute PDE residual via automatic differentiation
4. Loss = PDE_residual + BC_penalty + IC_penalty
5. Optimize with gradient-based methods

Advantages Over Standard PINNs/QPINNs

Metric	PINN	QPINN	GQPINN
Parameters	High	Medium	Low
Symmetry compliance	No	Partial	Guaranteed
Generalization	Baseline	Improved	Best
Training cost	High	Medium	Lower

Key Design Patterns

Pattern 1: Symmetry Detection → Circuit Construction

PDE Analysis → Symmetry Group → Generator Sets → Twirling → Ansatz

Pattern 2: Matched Training Protocols

When benchmarking GQPINNs against baselines:

• Use identical training point distributions
• Match optimization hyperparameters
• Compare at equal parameter counts
• Report mean absolute error (MAE) as primary metric

PDE Categories for Application

• Linear PDEs: Heat equation, wave equation, Schrödinger equation
• Nonlinear PDEs: Burgers equation, KdV equation, Navier-Stokes
• Geometric PDEs: Problems on manifolds with intrinsic symmetries

Implementation Notes

• Use parameterized quantum circuits (PQCs) as base ansatz
• Number of qubits scales with problem complexity
• Twirling can be approximated via random sampling for large groups
• For continuous groups, discretize to finite subgroup for practical twirling

References

• GQPINN paper: arxiv:2605.02352 (Tam, Safari, Matsuyama, 2026)
• Geometric Quantum Machine Learning: Meyer et al. (2022)
• Physics-Informed Neural Networks: Raissi et al. (2019)