GMM-PIELM
Accelerating Stiff PDE Resolution via Gaussian Mixture Model based Physics-Informed Extreme Learning Machines (PI-ELMs)
Partial Differential Equations (PDEs) serve as the bedrock for modeling physical phenomena across engineering domains. While classical discretization approaches like FEM, FDM, and FVM are mathematically rigorous, they suffer from intensive mesh-generation bottlenecks in complex 3D structures. Physics-Informed Neural Networks (PINNs) offer a mesh-free path forward, yet they remain bottlenecked by slow iterative training, an uninterpretable “black-box” architecture, and volatile hyperparameter sensitivity.
This project showcases RBF-PIELM(Srinivasan et al., 2026) and GMM-PIELM (Srinivasan & Srinivasan, 2026), which synergize the physics-constrained optimization of PINNs with the lightning-fast, non-iterative linear solving paradigm of Extreme Learning Machines (ELMs). By substituting global activation functions with localized Gaussian kernels, we enable physics-informed initialization and dynamic, error-driven resource allocation.
Core Architectural Framework
Standard PIELMs rely on physics-agnostic, random input parameters, leaving hidden features completely uninterpretable. The Radial Basis Function-based PIELM (RBF-PIELM) addresses this by anchoring Gaussian kernels directly inside the computational domain:
\[\phi_{i}(x;x_{0}^{(i)},\sigma_{i}) = \exp\left(-\frac{||x-x_{0}^{(i)}||^{2}}{2\sigma_{i}^{2}}\right)\]To resolve highly localized stiff dynamics where manual heuristics fail, the Gaussian Mixture Model Adaptive PIELM (GMM-PIELM) treats the unnormalized $\log(1+\text{residual})$ field as a spatial probability density function (PDF) of the error profile. An Expectation-Maximization (EM) algorithm dynamically drives RBF centers to cluster closely within critical high-gradient zones, such as thin boundary layers.
Fluid Mechanics Benchmarks: Lid-Driven Cavity & Backstep Flow
We benchmarked the RBF-PIELM framework against traditional PINNs on canonical incompressible fluid flow problems modeled via the stream function-vorticity (biharmonic) formulation:
\[\psi_{xxxx} + 2\psi_{xxyy} + \psi_{yyyy} = 0\]Numerical experiments demonstrate that our mesh-free solver computes fluid structures accurately while crushing the training overhead of traditional networks.
- Extreme Acceleration: RBF-PIELM achieves comparable accuracy while executing 350x to 658x faster than PINNs.
- Sustainable Computing: Our non-iterative linear solve delivers a 2,000x to 3,800x reduction in energy consumption, cutting carbon footprints down from grams to mere milligrams of $CO_2$ per run.
- Parametric Efficiency: Achieved lower physical residual error while using up to 13.2x fewer parameters.
Scaling to Complex Geometry & High-Dimensional Finance
Beyond 2D fluid benchmarks, we validated the geometric versatility of the solver on highly complex 3D Gyroidal Minimal Surfaces governed by a Poisson system:
\[\frac{\partial^2 u}{\partial x^2} + \frac{\partial^2 u}{\partial y^2} + \frac{\partial^2 u}{\partial z^2} = R(x,y,z)\]Despite the highly convoluted, interconnected microfluidic channels, RBF-PIELM evaluates the full 3D domain space in just 6.88 seconds.
Derivative Pricing via Black-Scholes and Heston-Hull-White Systems
In quantitative finance, the framework was adapted to circumvent the “curse of dimensionality” when evaluating complex multi-factor derivatives. In (Srinivasan et al., 2025), we modeled options under the high-dimensional Heston-Hull-White (HHW) framework, which introduces simultaneous stochastic volatility and stochastic short interest rates:
\[\frac{\partial V}{\partial t} + rS\frac{\partial V}{\partial S} + \kappa_v(\theta_v - v)\frac{\partial V}{\partial v} + \kappa_r(\theta_r - r)\frac{\partial V}{\partial r} + \dots - rV = 0\]Our approach maps combined stochastic fluctuations 32x faster than PINNs, allowing for near real-time financial derivative discounting and risk assessment.
Solving the Inverse Problem: Real-Time Parameter Calibration
For practical market integration, a quantitative model must undergo fast calibration—recovering unobservable parameters (e.g., long-term volatility mean, mean-reversion speeds) from noisy market prices. Because optimization loops solve the forward PDE thousands of times, iterative neural methods are too slow for production settings.
By implementing a Teacher-Student validation protocol, we paired our fast linear solver with a Bayesian Optimization loop. Synthetic pricing surfaces generated by a high-capacity teacher network were intentionally corrupted with high-variance multiplicative noise ($\eta = 10\%$) to simulate chaotic real-world spreads.
Our framework accurately reconstructs the underlying price manifold while keeping calibration parameter errors under safe margins, proving its viability for real-time risk evaluation workflows.
References
2026
- Towards Sustainable Scientific Machine Learning: Fast and interpretable PDE Solvers via RBF-PIELMIn Proceedings of the 13th ACM IKDD International Conference on Data Science and Management of Data, 2026
- Learning Where the Physics Is: Probabilistic Adaptive Sampling for Stiff PDEsIn Workshop on Artificial Intelligence and Partial Differential Equations, 2026
2025
- Towards Fast Option Pricing PDE Solvers Powered by PIELMIn AI Meets Quantitative Finance: Stochastic Methods as a Two-Way Bridge Workshop, 2025