LieFlow: Symmetry & Dynamic Simulation

• LieFlow is a machine learning framework that integrates Lie group theory, flow matching, and geometric priors to automatically discover both continuous and discrete symmetries in data.
• It employs techniques such as exponential mapping and power-law time rescheduling to ensure precise symmetry sampling and robust dynamic scene modeling.
• Empirical validations demonstrate LieFlow’s superior performance in symmetry discovery, dynamic field modeling, and accelerated atomistic simulation compared to traditional methods.

LieFlow refers to a distinct class of machine learning frameworks leveraging Lie groups, flow-matching objectives, and geometric priors to address problems in symmetry discovery, dynamic scene modeling, and accelerated atomistic simulation. Multiple methodologies—each with this name or its variant—have been proposed, with notable instantiations in (1) symmetry discovery in data (Park et al., 23 Dec 2025), (2) physically consistent video-based dynamic field modeling (Qiao et al., 25 Feb 2026), and (3) accelerated simulation of atomic transport in materials (LiFlow) (Nam et al., 2024). Each utilizes core concepts of Lie group theory and flow-based learning but targets distinct domains. Their mathematical and algorithmic structures, as well as empirical performance, are detailed below for technically sophisticated audiences.

1. Symmetry Discovery with LieFlow: Flow Matching on Lie Groups

LieFlow (Park et al., 23 Dec 2025) frames the automatic discovery of symmetries as a flow matching problem on Lie groups. Here, a hypothesis group $G$ (e.g., $\mathrm{SO}(n)$, $\mathrm{GL}(d)$) is assumed to contain the unknown stabilizer subgroup $H = \{ h \in G \mid q(h\cdot x) = q(x) \}$, where $q$ is the (unknown) data distribution and the action $h\cdot x$ is known.

Mathematical Formulation and Training Objective

The method generalizes conditional flow matching from Euclidean space to Lie groups. For data points $x_1 \sim q$ and group elements $g \sim p(G)$, construct the initial point as $x_0 = g \cdot x_1$. The interpolation path on $G$ is parameterized via the exponential map: $\mathrm{SO}(n)$0 with $\mathrm{SO}(n)$1 the relevant Lie algebra element. The time-evolved point is

$\mathrm{SO}(n)$2

with target velocity $\mathrm{SO}(n)$3, constant in $\mathrm{SO}(n)$4. The regression objective for the neural vector field $\mathrm{SO}(n)$5 is

$\mathrm{SO}(n)$6

using a left-invariant Riemannian metric $\mathrm{SO}(n)$7 on $\mathrm{SO}(n)$8.

Distribution Learning over $\mathrm{SO}(n)$9

By pushing a simple prior $\mathrm{GL}(d)$0 through the learned flow, LieFlow infers a distribution $\mathrm{GL}(d)$1 supported on the true symmetry subgroup $\mathrm{GL}(d)$2: continuous $\mathrm{GL}(d)$3 yields spread distributions; discrete $\mathrm{GL}(d)$4 yields multimodal ones. After training, sampling $\mathrm{GL}(d)$5 from $\mathrm{GL}(d)$6 and applying the learned dynamical flow yields a group element $\mathrm{GL}(d)$7.

2. Algorithmic Structure and Practical Implementation

A typical workflow consists of:

1. Sampling $\mathrm{GL}(d)$8 and $\mathrm{GL}(d)$9, then setting $H = \{ h \in G \mid q(h\cdot x) = q(x) \}$0, $H = \{ h \in G \mid q(h\cdot x) = q(x) \}$1.
2. For $H = \{ h \in G \mid q(h\cdot x) = q(x) \}$2, generating $H = \{ h \in G \mid q(h\cdot x) = q(x) \}$3.
3. Computing a gradient update for $H = \{ h \in G \mid q(h\cdot x) = q(x) \}$4 by minimizing $H = \{ h \in G \mid q(h\cdot x) = q(x) \}$5.
4. For generation and symmetry sampling, integrating $H = \{ h \in G \mid q(h\cdot x) = q(x) \}$6 in small steps and accumulating transforms to retrieve $H = \{ h \in G \mid q(h\cdot x) = q(x) \}$7.

Architecturally, a compact MLP typically suffices for $H = \{ h \in G \mid q(h\cdot x) = q(x) \}$8. To capture reflection symmetries (e.g., dihedral $H = \{ h \in G \mid q(h\cdot x) = q(x) \}$9), $q$0 is embedded into $q$1, and networks over $q$2 are used.

3. Addressing Last-Minute Convergence and Time Scheduling

In discovery of discrete symmetry groups, "last-minute convergence" arises: for much of $q$3, the posterior $q$4 is nearly uniform and target velocity is near zero, leading to abrupt convergence at $q$5. To mitigate, LieFlow samples time from a power-law $q$6 ($q$7), concentrating updates near $q$8 and promoting early resolution of group modes (Park et al., 23 Dec 2025).

4. Empirical Validation and Quantitative Results

LieFlow is validated on 2D and 3D synthetic point clouds with both finite ($q$9, $h\cdot x$0, tetrahedral, octahedral) and continuous symmetry groups. Representative performances include:

• $h\cdot x$1D: Wasserstein-1 distance $h\cdot x$2 (SO(2)→$h\cdot x$3), vs. $h\cdot x$4 for LieGAN baseline.
• $h\cdot x$5D: Wasserstein-1 $h\cdot x$6 on tetrahedral and octahedral groups, vs. $h\cdot x$7 for LieGAN.
• Multi-object and approximate-symmetry robustness are demonstrated.
• Without power-law time rescheduling, only small or simple finite groups are reliably discovered; with it ($h\cdot x$8), higher-order discrete symmetries (octahedral, $h\cdot x$9 elements) are correctly identified, but icosahedral ($x_1 \sim q$0 elements) remains challenging.

5. Limitations and Future Directions

Limitations include persistent last-minute convergence in high-order discrete groups, difficulties in scaling to high-dimensional or non-compact hypothesis groups, and the challenge of only approximately symmetric real-world datasets. Proposed future directions are:

• Riemannian diffusion or score-matching formalisms to avoid discontinuities at cut loci.
• Group-aware neural architectures, specifically equivariant models, for enhanced learning efficiency.
• Downstream integration of discovered symmetries into equivariant prediction pipelines.
• Extending the formalism to approximately symmetric cases by matching distributions over the observed symmetry orbits.

6. Related Approaches in Lie-Group Flow Modeling

While "LieFlow" in this context (Park et al., 23 Dec 2025) is symmetry-focused, multiple approaches in the literature harness Lie group flow matching for diverse purposes:

• Video Dynamic Fields Modeling: "Lie Flow: Video Dynamic Fields Modeling and Predicting with Lie Algebra as Geometric Physics Principle" (Qiao et al., 25 Feb 2026) explicitly parameterizes a spatio-temporal transformation field using the Lie algebra $x_1 \sim q$1 of the Special Euclidean group, enforcing collective translation and rotation priors to guarantee physically consistent 4D reconstructions.
• Accelerated Atomic Simulation: "LiFlow" (Nam et al., 2024) learns a conditional displacement model for atomic configurations using an optimal transport-inspired flow-matching loss between Gaussian priors and empirical MD displacements, enabling orders of magnitude acceleration over standard molecular dynamics.

The table below summarizes these LieFlow variants:

Variant	Domain	Core Mechanism
LieFlow (Park et al., 23 Dec 2025)	Symmetry discovery	Flow matching on Lie group manifolds
Lie Flow (Qiao et al., 25 Feb 2026)	Dynamic scene modeling	SE(3) Lie algebra field integration
LiFlow (Nam et al., 2024)	Atomistic simulation	OT-based flow on displacement space

All employ neural flow fields on group or vector spaces, with loss functions structured to enforce group-theoretic or physical priors.

7. Broader Impact and Theoretical Significance

The unifying feature of LieFlow-based methodologies is the encoding of group structure directly within the learning dynamics. This enables:

• Unsupervised discovery of both continuous and discrete symmetry groups from unlabelled data.
• Embedding of rigid-body and articulated motion priors for physically consistent spatiotemporal modeling.
• Substantial acceleration and improved scaling of high-fidelity simulation in physical sciences.

The approach bridges optimal transport, equivariant learning, and geometric deep learning, establishing a mathematically principled framework for inferring and leveraging symmetry and structure beyond traditional supervised or model-driven pipelines. For researchers, LieFlow provides both a theoretical tool for symmetry analysis and a practical foundation for next-generation data-efficient and physically faithful models (Park et al., 23 Dec 2025, Qiao et al., 25 Feb 2026, Nam et al., 2024).