[Paper Review] A composable autoencoder-based iterative algorithm for accelerating numerical simulations

Rishikesh Ranade, Chris Hill+4 more|arXiv (Cornell University)|Oct 7, 2021

Model Reduction and Neural Networks63 references4 citations

TL;DR

This paper proposes CoAE-MLSim, a composable autoencoder-based iterative machine learning framework that accelerates numerical simulations of partial differential equations (PDEs) by learning low-dimensional, local representations on subdomains and enforcing flux conservation across them. The method achieves high accuracy and generalization across diverse PDE conditions—outperforming FNO and U-Net in error accumulation—while enabling unsupervised, scalable inference with minimal data and seamless coupling to traditional solvers.

ABSTRACT

Numerical simulations for engineering applications solve partial differential equations (PDE) to model various physical processes. Traditional PDE solvers are very accurate but computationally costly. On the other hand, Machine Learning (ML) methods offer a significant computational speedup but face challenges with accuracy and generalization to different PDE conditions, such as geometry, boundary conditions, initial conditions and PDE source terms. In this work, we propose a novel ML-based approach, CoAE-MLSim (Composable AutoEncoder Machine Learning Simulation), which is an unsupervised, lower-dimensional, local method, that is motivated from key ideas used in commercial PDE solvers. This allows our approach to learn better with relatively fewer samples of PDE solutions. The proposed ML-approach is compared against commercial solvers for better benchmarks as well as latest ML-approaches for solving PDEs. It is tested for a variety of complex engineering cases to demonstrate its computational speed, accuracy, scalability, and generalization across different PDE conditions. The results show that our approach captures physics accurately across all metrics of comparison (including measures such as results on section cuts and lines).

Motivation & Objective

• Address the computational cost of traditional PDE solvers while improving generalization and accuracy over existing ML-based PDE solvers.
• Overcome limitations of static, data-driven ML models that struggle with sparse, high-dimensional PDE conditions such as varying geometry, boundary conditions, and source terms.
• Integrate time-tested PDE solver principles—like domain decomposition, local subdomain processing, and iterative consistency—into a learnable, composable ML framework.
• Enable unsupervised, low-data, and scalable inference by training autoencoders on local subdomains and using iterative algorithms to enforce global consistency.
• Demonstrate strong performance across diverse engineering problems, including transient flows and vortex decay, with metrics aligned to commercial solvers.

Proposed method

• The method operates on local 3D subdomains (e.g., 8×8×8 or 16×16×16 voxels), learning compressed representations of PDE solutions using solution autoencoders.
• Flux conservation between neighboring subdomains is modeled via flux conservation autoencoders trained on stacked latent vectors of adjacent subdomains.
• PDE condition autoencoders (for geometry, boundary conditions, and source terms) are trained independently using random or sparse samples, enabling generalization.
• An iterative inference algorithm combines pretrained autoencoders to enforce local consistency across the full computational domain, mimicking traditional solver convergence.
• The framework supports unsupervised training and can be coupled with traditional PDE solvers to accelerate convergence and improve accuracy.
• Network architectures include a U-Net-like encoder-decoder with skip connections and max-pooling/bilinear upsampling; total parameters: ~7.4M for solution autoencoder, ~465k for FNO baseline.

Experimental results

Research questions

• RQ1 Can a machine learning framework that integrates core principles from traditional PDE solvers achieve better accuracy and generalization than end-to-end ML models on diverse PDE conditions?
• RQ2 To what extent can local, low-dimensional autoencoder representations on subdomains improve data efficiency and scalability in PDE simulation?
• RQ3 How well does the iterative inferencing strategy with composable autoencoders preserve physical consistency and reduce error accumulation over time in transient problems?
• RQ4 Can the proposed method generalize across high-dimensional, sparse PDE conditions (e.g., varying Reynolds numbers, complex geometries) without retraining?
• RQ5 How does the performance of CoAE-MLSim compare to state-of-the-art ML models like FNO and U-Net in terms of error accumulation and physical fidelity?

Key findings

• CoAE-MLSim achieves high accuracy in predicting vortex decay and transient flow over a cylinder, with contour comparisons closely matching Ansys Fluent results.
• Error accumulation in the vortex decay case remains acceptable and significantly lower than FNO and U-Net across time steps, indicating superior long-term stability.
• The method successfully captures the vortex shedding frequency and total drag force evolution over time, with predictions closely aligned with commercial solver benchmarks.
• The framework generalizes well across diverse PDE conditions, including changes in Reynolds number, geometry, and boundary conditions, due to its local, composable, and unsupervised design.
• Training requires only 100–1000 solution samples, and the model achieves strong performance with minimal parameters (7.4M for solution autoencoder), demonstrating data efficiency.
• The iterative inferencing strategy enables seamless coupling with traditional solvers, enhancing convergence and accuracy without sacrificing speed or generalization.