Deep Energy Method with Large Language Model assistance: an open-source Streamlit-based platform for solving variational PDEs

Physics-informed neural networks (PINNs) in energy form, also known as the deep energy method (DEM), offer advantages over strong-form PINNs such as lower-order derivatives and fewer hyperparameters, yet dedicated and user-friendly software for energy-form PINNs remains scarce. To address this gap, we present \textbf{LM-DEM} (Large-Model-assisted Deep Energy Method), an open-source, Streamlit-based platform for solving variational partial differential equations (PDEs) in computational mechanics. LM-DEM integrates large language models (LLMs) for geometry modeling: users can generate Gmsh-compatible geometries directly from natural language descriptions or images, significantly reducing the burden of traditional geometry preprocessing. The solution process is driven by the deep energy method, while finite element solutions can be obtained in parallel. The framework supports built-in problems including Poisson, screened Poisson, linear elasticity, and hyperelasticity in two and three dimensions, as well as user-defined energy functionals analogous to the \texttt{UMAT} interface in Abaqus. The source code is available at https://github.com/yizheng-wang/LMDEM, and a web-based version is accessible at https://ai4m.llmdem.com. LM-DEM aims to lower the barrier for practitioners and beginners to adopt energy-form PINNs for variational PDE problems.

💡 Research Summary

The paper introduces LM‑DEM, an open‑source, Streamlit‑based platform that combines the Deep Energy Method (DEM) with large language model (LLM) assistance to solve variational partial differential equations (PDEs) in computational mechanics. The authors begin by reviewing the landscape of AI‑for‑PDEs, noting that most existing PINN libraries focus on the strong‑form formulation, which requires higher‑order derivatives and multiple penalty hyper‑parameters, often leading to unstable training and cumbersome tuning. In contrast, the energy‑form (or variational) PINN, known as DEM, only needs first‑order derivatives and a single energy functional, offering better numerical stability, accuracy, and computational efficiency.

LM‑DEM addresses the lack of user‑friendly software for energy‑form PINNs. Its workflow consists of three stages: (1) preprocessing, where geometry is generated automatically from natural‑language prompts or images via an LLM (default OpenAI GPT‑4o, with alternatives such as DeepSeek, Anthropic, Google, or local Ollama). The LLM outputs a Gmsh‑compatible *.geo script; error‑feedback loops enable iterative refinement, especially for 3D models, achieving a reported success rate above 90 %. Users can also import their own *.msh files. (2) solver configuration, where the UI lets users set mesh resolution, element type, quadrature rule, derivative evaluation method, material model (Poisson, screened Poisson, linear elasticity, hyperelasticity, or custom), boundary conditions, neural‑network architecture (layers, activation, width), and training hyper‑parameters (epochs, batch size, learning rate, CUDA). Dirichlet conditions are enforced a priori using a split form u(x)=u_p(x)+D(x)·u_g(x), where u_p satisfies the prescribed displacement, D(x) is a radial‑basis‑function distance field, and u_g is the trainable network. The loss is the total potential energy (domain strain energy minus work of Neumann tractions and body forces), evaluated by Monte‑Carlo sampling or higher‑order quadrature. Only the parameters of u_g are optimized. (3) post‑processing, where the DEM solution is visualized and compared against a parallel finite‑element solution computed on the same mesh, providing a reference benchmark.

The platform ships with built‑in 2D and 3D examples for Poisson and elasticity problems, and it supports user‑defined energy density functions through an interface analogous to Abaqus UMAT, enabling extensions to J2 plasticity, phase‑field fracture, or any custom constitutive law. The code is written in Python, leveraging PyTorch/JAX for automatic differentiation and GPU acceleration; the Streamlit front‑end offers a web version (CPU‑only) and a local version (GPU‑enabled). Experimental results demonstrate that LM‑DEM reproduces FEM solutions with errors typically below 2 % for linear and nonlinear problems, while training times are comparable to or better than strong‑form PINNs because of the reduced derivative order and fewer hyper‑parameters. The LLM‑driven geometry creation dramatically cuts preprocessing time, turning a task that traditionally requires hours of manual CAD work into a few seconds of prompt‑based generation.

Limitations include longer response times for complex 3D geometry generation due to multiple LLM calls, dependence on prompt quality, and the fact that DEM is naturally suited only to problems that admit a variational (energy) formulation; non‑variational PDEs (e.g., certain convection‑dominated flows) would still need strong‑form PINNs or alternative weak‑form strategies. The web interface currently lacks Gmsh GUI editing and GPU support, which are only available in the locally deployed version.

In conclusion, LM‑DEM provides a comprehensive, beginner‑friendly environment that lowers the entry barrier to energy‑form PINNs, integrates modern LLM capabilities for geometry preprocessing, and offers parallel FEM validation. Future work will focus on accelerating 3D geometry synthesis, extending support to non‑variational PDEs, adding GPU acceleration to the web service, and developing systematic prompt‑engineering guidelines. Overall, LM‑DEM represents a significant step toward practical, AI‑augmented computational mechanics.