A Particle-Field Algorithm with Neural Interpolation for a Parabolic-Hyperbolic Chemotaxis System in 3D

Basic Information

• Paper ID: 2510.13199
• Title: A Particle-Field Algorithm with Neural Interpolation for a Parabolic-Hyperbolic Chemotaxis System in 3D
• Authors: Jongwon David Kim, Jack Xin (University of California, Irvine)
• Classification: math.NA cs.NA
• Publication Date: October 15, 2025 (arXiv preprint)
• Paper Link: https://arxiv.org/abs/2510.13199v1

Abstract

Tumor angiogenesis involves the process of tumor cells migrating toward blood vessels to acquire nutrients. Angiogenesis and general chemotaxis systems have been modeled through partial differential equations (PDEs), necessitating numerical methods for approximate solutions. This paper investigates the parabolic-hyperbolic Keller-Segel (PHKS) system in three-dimensional space, a model derived from angiogenesis literature. To solve the PHKS system, the authors develop the Neural Stochastic Interacting Particle Field (NSIPF) method, wherein the density variable is represented as an empirical measure of particles, and the field variable (chemoattractant concentration) is approximated through a convolutional neural network (CNN). The paper discusses the performance of NSIPF in computing multimodal solutions of the system.

Research Background and Motivation

1. Problems to be Addressed:
   • Mathematical modeling and numerical solution of cell migration toward blood vessels in tumor angiogenesis
   • Efficient numerical computation of parabolic-hyperbolic Keller-Segel systems in three-dimensional space
2. Problem Significance:
   • Angiogenesis is a critical biological process in tumor metastasis; understanding its mechanisms is important for cancer treatment
   • Chemotaxis systems play important roles in multiple biological processes including bacterial foraging and tissue formation
3. Limitations of Existing Methods:
   • Traditional finite difference methods (FDM) are time-consuming for high-resolution three-dimensional computations
   • Classical spline interpolation methods are complex to implement in three-dimensional space and computationally expensive
   • Lack of efficient particle methods specifically designed for parabolic-hyperbolic systems
4. Research Motivation:
   • Leveraging neural networks' high-dimensional approximation capabilities to improve interpolation efficiency
   • Combining the mesh-free advantages of particle methods with CNN interpolation performance
   • Providing more efficient numerical solution schemes for three-dimensional chemotaxis systems

Core Contributions

1. Proposed the Neural Stochastic Interacting Particle Field (NSIPF) method, integrating CNN interpolators into stochastic particle-field algorithms
2. Established theoretical foundations of chaos propagation for the PHKS system, providing theoretical support for neural interpolation through moderate interactions
3. Developed efficient CNN architectures, successfully implementing interpolation for 2D and 3D systems using radially symmetric solutions as training data
4. Achieved significant computational efficiency improvements, delivering orders of magnitude speedup compared to traditional methods while maintaining solution quality

Methodology Details

Problem Formulation

Solving the parabolic-hyperbolic Keller-Segel system in three-dimensional space:

ρₜ = ∇ · (γ∇ρ - χρ∇c)  (1)
cₜ = -cρ                  (2)

where ρ is bacterial density, c is chemoattractant concentration, γ is the diffusion coefficient, and χ is the chemotactic sensitivity.

Model Architecture

1. Particle Representation

The density function is approximated through the empirical measure of particles:

ρₜ ≈ (M₀/P) Σⱼ₌₁ᴾ δ(x - Xⱼₜ)

where P is the number of particles and M₀ is the conserved mass.

2. SDE Formulation

Particle motion follows stochastic differential equations:

dXᵢₜ = χ∇c(Xᵢₜ,t)dt + √(2γ)dBᵢₜ

3. CNN Interpolator Architecture

• Input Layer: Low-resolution concentration field data
• Convolutional Layers: 5 layers of 3D convolution with channel progression 1→16→32→32→32→16→1
• Activation Function: ReLU
• Skip Connections: Enhancing gradient propagation
• Output: High-resolution interpolated field

4. NSIPF Algorithm Flow

Algorithm 1: Neural SIPF
1. Initialize particle positions based on ρ₀
2. For each time step:
   a. Bin particles to obtain ρₙ
   b. Update concentration: cⁿ⁺¹ = cⁿ - δt·cⁿ·ρⁿ
   c. CNN interpolation to obtain c(x,tₙ)
   d. Compute ∇c and update particle positions

Technical Innovations

1. Neural Interpolation Replacing Classical Interpolation:
   • Using CNN instead of spline interpolation to compute concentration field gradients
   • Training data derived from inexpensive solutions of one-dimensional radial systems
   • Supporting generalization to non-radial initial conditions
2. Chaos Propagation Theory Foundation:
   • Establishing theoretical basis through mollifier regularization
   • Moderate interactions ensuring convergence of particle systems
   • Providing mathematical justification for CNN interpolation
3. Efficient Training Strategy:
   • Data augmentation using radial solutions (downsampling, translation, blurring)
   • Patch-based training reducing 3D computational burden
   • Achieving good interpolation effects within 100 training epochs

Experimental Setup

Dataset

• Training Data: 50 solutions of radial PHKS systems computed via finite difference methods
• Data Augmentation: Patch-based downsampling, translation, and blurring of solutions
• Test Cases:
  • Single Gaussian blob diffusion
  • Dual blob aggregation
  • Ring boundary aggregation

Evaluation Metrics

• Relative L² Error: √(Σ(f_num - f_ref)²)/√(Σf_ref²)
• Runtime: Computational time at different resolutions
• Convergence: Convergence rates with respect to particle number P and time step δt

Comparison Methods

• Finite Difference Method (FDM): Traditional grid-based method
• Classical SIPF: Particle method using spline interpolation
• Neural SIPF: Proposed CNN interpolation method

Implementation Details

• Optimizer: Adam with learning rate 10⁻³
• Batch Size: 4
• Loss Function: Mean Squared Error (MSE)
• Training Duration: 9753.73 seconds (NVIDIA GTX 1080)
• Boundary Conditions: Neumann boundary conditions

Experimental Results

Main Results

Runtime Comparison (3D, P=20k particles)

Particle Number Scalability

Convergence Analysis

• Particle Number Convergence: e(P) = O(P^(-0.45)), approaching theoretical expectation of O(P^(-1/2))
• Time Step Convergence: e(δt) = O(δt^0.93), approaching first-order accuracy

Case Studies

1. Single Blob Diffusion: Neural SIPF successfully captures diffusion behavior, consistent with FDM results
2. Dual Blob Aggregation: Two eccentric blobs aggregate toward the central food source, demonstrating correct aggregation dynamics
3. Ring Boundary: Particles aggregate along the ring-shaped food source boundary, validating method adaptability to non-blob initial conditions

Experimental Findings

• Although CNN interpolation is an approximate method (training loss does not fully converge to zero), it effectively captures system dynamics
• Neural SIPF achieves orders of magnitude speedup compared to traditional methods in high-resolution 3D computations
• The method demonstrates good generalization to data types outside the training set (e.g., ring-shaped initial conditions)

Related Work

Main Research Directions in the Field

1. Theoretical Analysis of Keller-Segel Systems: Existence, uniqueness, and blow-up phenomena of solutions
2. Development of Numerical Methods: Finite difference, finite element, and particle methods
3. Biological Applications: Bacterial chemotaxis, tumor growth, and tissue formation models

Relationship to Related Work

• Based on Wang et al.'s particle methods for parabolic-parabolic KS systems 9
• Extending Hu et al.'s particle algorithms for tactile systems 4
• Drawing from neural interpolation techniques in meteorology 7,8

Advantages Over Related Work

• First application of CNN interpolation to parabolic-hyperbolic chemotaxis systems
• Providing complete theoretical foundation of chaos propagation
• Achieving significant breakthrough in efficient three-dimensional computation

Conclusions and Discussion

Main Conclusions

1. The NSIPF method successfully combines the mesh-free advantages of particle methods with CNN's high-dimensional interpolation capabilities
2. Chaos propagation theory provides theoretical foundation for applying neural interpolation to PDE numerical solutions
3. The method achieves significant computational efficiency improvements while maintaining solution quality

Limitations

1. Training Data Dependency: Requires pre-computed radial solutions as training data
2. Interpolation Accuracy: CNN interpolation is an approximate method with inherent errors
3. Parameter Tuning: Neural network hyperparameters require problem-specific adjustment
4. Theoretical Completeness: Chaos propagation theory for non-smooth initial conditions c₀ remains incomplete

Future Directions

1. Improved Interpolation Performance: Enhancing runtime and interpolation accuracy
2. Generative AI: Training generative models on particle solutions, completely bypassing mechanistic model computation
3. Extended Applications: Generalizing the method to more complex biological systems and other PDE types

In-Depth Evaluation

Strengths

1. Strong Method Innovation: First systematic application of CNN interpolation to chemotaxis system numerical solutions
2. Solid Theoretical Foundation: Chaos propagation theory provides rigorous mathematical basis for the method
3. Comprehensive Experiments: Covering multiple initial conditions with detailed performance comparisons
4. High Practical Value: Achieving significant efficiency improvements in high-resolution 3D computations
5. Clear Presentation: Detailed algorithm descriptions and complete implementation details

Weaknesses

1. Training Cost: Although faster than traditional methods, pre-training CNN adds upfront costs
2. Generalization Capability: Generalization to complex initial conditions outside training data requires further verification
3. Error Analysis: Theoretical analysis of CNN interpolation errors lacks depth
4. Parameter Sensitivity: Insufficient discussion of method sensitivity to hyperparameter choices

Impact

1. Academic Contribution: Provides new paradigm for combining PDE numerical methods with machine learning
2. Application Prospects: Broad applications in computational biology, materials science, and other fields requiring efficient PDE solutions
3. Reproducibility: Detailed algorithm descriptions and relatively simple implementation facilitate reproduction and dissemination

Applicable Scenarios

1. Three-Dimensional Chemotaxis Systems: Particularly suitable for large-scale 3D computations requiring long-time evolution
2. Multiscale Problems: Particle methods naturally suited for handling multiscale phenomena
3. Real-Time Simulation: High efficiency enables applications requiring rapid response
4. Parameter Studies: Facilitates extensive parameter sweeps and sensitivity analyses

References

The paper cites 18 important references, primarily including:

• Keller & Segel (1970): Foundational work on chemotaxis systems
• Corrias et al. (2003): Theoretical analysis of PHKS systems
• Chaintron & Diez (2022): Survey of chaos propagation theory
• Wang et al. (2025): Related particle method research

Overall Assessment: This is a high-quality numerical analysis paper that successfully applies deep learning techniques to PDE numerical solutions, combining solid theoretical foundations with significant practical value. The method demonstrates strong innovation, comprehensive experimental validation, and important implications for advancing the interdisciplinary field of computational mathematics and machine learning.