Boosted Physics-Informed Neural Networks for Solving Homogeneous and Non-Homogeneous Lane-Emden Type Equations

Luke Oluwaseye JoelarXiv 2026 · arXiv

TL;DR

This paper introduces a hybrid framework that marries Physics-Informed Neural Networks (PINNs) with ensemble boosting methods like Random Forest and LightGBM. By using machine learning to model and correct the residual errors of a standard PINN, the authors achieve high-fidelity solutions for complex Lane-Emden type differential equations, outperforming baseline PINNs by up to three orders of magnitude.

What Problem Does This Paper Address?

The authors focus on solving Lane-Emden type equations, a class of nonlinear ordinary differential equations (ODEs) fundamental to astrophysics for modeling stellar structure, density, and thermal behavior in gas spheres. These equations are notoriously difficult to solve because they are nonlinear and often contain singularities (e.g., terms like $1/x$ that blow up at the origin).

While Physics-Informed Neural Networks (PINNs) have emerged as a powerful tool for solving such equations without needing traditional discretization, they often hit a "performance ceiling." Optimization challenges, such as getting stuck in local minima, and approximation errors can limit their accuracy. The paper seeks to break through this ceiling by introducing a data-driven error-correction stage.

Methods

The proposed "Boosted PINN" framework operates in a systematic two-stage process:

  1. PINN Stage: A standard fully connected neural network is trained to approximate the solution $\theta(\xi)$. The loss function encodes the physics of the problem, including the ODE residual and initial/boundary conditions. This provides a physics-constrained baseline solution.
  2. Boosting Stage: The authors calculate the residual error—the difference between the PINN's prediction and the exact solution. They then train secondary machine learning models—specifically LightGBM, Random Forest (RF), XGBoost, and Support Vector Regression (SVR)—to learn the mapping from spatial coordinates to these error values.

The final, high-fidelity prediction is generated by adding the predicted error correction back to the original PINN solution. The authors used Bayesian optimization (via the Optuna framework) to fine-tune the hyperparameters of the underlying PINNs for each specific problem.

Figure 1: Boosted PINNs Flowchart

Datasets

The study validates the framework using six benchmark problems involving different variations of the Lane-Emden equation (both homogeneous and non-homogeneous).

  • These problems are sourced from established literature in mathematical physics and astrophysics.
  • The "dataset" for the boosting stage is generated internally: it consists of spatial coordinates $x$ as features and the calculated PINN residual errors $r(x)$ as targets.
  • Performance is measured against known exact solutions for these benchmarks.

Key Insights and Ideas

  • Decoupled Refinement: A key advantage of this approach is that it decouples physics-based approximation from data-driven refinement. The ML models correct the errors without needing to alter the complex physics formulation inside the PINN's loss function.
  • Tree-Based Models Excel: Interestingly, tree-based ensemble methods like Random Forest and LightGBM consistently provided the best error corrections. This suggests that PINN residual errors often contain non-linear patterns that bagging and boosting algorithms are particularly well-suited to capture.
  • Handling Singularities: Because PINNs are mesh-free, they handle the singularities at $x=0$ in Lane-Emden equations more robustly than traditional grid-based numerical methods.

Key Results

The results across all six problems demonstrate that boosting significantly enhances accuracy:

  • Significant Error Reduction: In many cases, the boosted models reduced the absolute error by one to three orders of magnitude compared to the standard PINN. For example, in Problem 1, the standard PINN error was $\sim 10^{-5}$, while the RF-boosted version achieved errors as low as $\sim 10^{-8}$.
  • Superiority Over Existing Methods: The framework outperformed several traditional and hybrid methods in the literature, including Chebyshev Neural Networks (ChNN), Multilayer Perceptrons (MLP), and Monte Carlo-based approaches.
  • Stability: PINN errors remained relatively stable across the domain, whereas errors in traditional methods like PS-SQP often increased by orders of magnitude as they approached the boundary ($x \to 1$).
Figure 2: Comparison of absolute errors for Problem 1, showing the dramatic improvement of boosted models over the original PINN.

Limitations and Future Work

  • Requirement of Exact Solutions: A primary limitation is that the current boosting stage requires the "exact" solution to calculate residuals for training. In practical scenarios where the solution is unknown, the authors would need to investigate using the physics-based residual (the ODE itself) as the target for the boosting models.
  • Scope of Equations: This study is limited to Ordinary Differential Equations (ODEs). The authors note that extending this framework to Partial Differential Equations (PDEs), which are more prevalent in complex scientific modeling, is a critical next step.
  • Computational Overhead: Adding a second stage of training for multiple ML models increases the total computational cost, though the authors argue the gain in fidelity justifies the overhead.

Summary updated

Abstract

The Lane-Emden Equation is a class of nonlinear differential equations commonly used to model phenomena in astrophysics and mathematical physics, such as stellar structure and thermal behavior in polytropic gas spheres. This study presents a novel hybrid framework that integrates Physics-Informed Neural Networks (PINNs) with boosting methods to enhance the accuracy and robustness of solutions for Lane-Emden type differential equations. Standard PINNs leverage physical laws encoded within the loss function, but their predictive performance can be constrained by optimization challenges and approximation errors. To address this, we introduce a post-processing boosting stage where the residual errors from the PINN prediction are modeled using some regression algorithms - including LightGBM, Random Forest, XGBoost, and Support Vector Regression (SVR). The outputs from these boosted models are aggregated to generate a corrected, high-fidelity prediction. The proposed methodology is systematically validated through quantitative metrics and comparative analysis against the standard PINN baseline. The results demonstrate that the boosted PINN framework significantly reduces prediction error, and improves generalization, offering a powerful and flexible approach for solving scientific and engineering problems governed by differential equations.
BibTeX
@misc{joel2026boosted,
  author = {Luke Oluwaseye Joel},
  title = {{Boosted Physics-Informed Neural Networks for Solving Homogeneous and Non-Homogeneous Lane-Emden Type Equations}},
  year = {2026},
  eprint = {2611.00701},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.00701},
}
Figure