General Explicit Network (GEN): A novel deep learning architecture for solving partial differential equations

A team of researchers including Genwei Ma has introduced the General Explicit Network (GEN), a new architecture designed to solve partial differential equations (PDEs) more effectively than current methods. The work directly addresses the limitations of Physics-Informed Neural Networks (PINNs), which have struggled to move beyond academic research. PINNs typically use a 'point-to-point' fitting approach with continuous activation functions, which can capture local solution characteristics but often fail to produce robust, extensible results for complex real-world problems.

The GEN architecture proposes a fundamental shift to a 'point-to-function' paradigm for PDE solving. Instead of fitting discrete data points, the model constructs a 'function' component based on prior knowledge of the original PDEs, utilizing corresponding basis functions. This method leverages known mathematical structure, allowing the network to learn more generalizable solutions. The authors' experimental results demonstrate that this approach enables the discovery of solutions with high robustness and strong extensibility, key requirements for practical applications in fields like medical physics and engineering.

This development is significant because it tackles the core issue of poor generalizability in neural PDE solvers. By incorporating domain knowledge directly into the network's functional construction, GEN moves away from purely data-driven interpolation. This could unlock more reliable use of machine learning for simulating physical systems, fluid dynamics, and material science, where accurate and stable predictions are critical.