Physics-informed neural networks with hard constraints for inverse design

Abstract: Inverse design arises in a variety of areas in engineering such as acoustic, mechanics, thermal/electronic transport, electromagnetism, and optics. Topology optimization is a major form of inverse design, where we optimize a designed geometry to achieve targeted properties and the geometry is parameterized by a density function. This optimization is challenging, because it has a very high dimensionality and is usually constrained by partial differential equations (PDEs) and additional inequalities. Here, we propose a new deep learning method -- physics-informed neural networks with hard constraints (hPINNs) -- for solving topology optimization. hPINN leverages the recent development of PINNs for solving PDEs, and thus does not rely on any numerical PDE solver. However, all the constraints in PINNs are soft constraints, and hence we impose hard constraints by using the penalty method and the augmented Lagrangian method. We demonstrate the effectiveness of hPINN for a holography problem in optics and a fluid problem of Stokes flow. We achieve the same objective as conventional PDE-constrained optimization methods based on adjoint methods and numerical PDE solvers, but find that the design obtained from hPINN is often simpler and smoother for problems whose solution is not unique. Moreover, the implementation of inverse design with hPINN can be easier than that of conventional methods.

• The paper introduces a hard-constrained PINN framework that directly incorporates Maxwell's equations for precise inverse design.
• It employs FDFD and FEM methods with MMA optimization, achieving robust convergence within 5000 iterations for electromagnetic permittivity.
• The approach pioneers integrating physical constraints into AI-driven PDE solvers, promising advances in engineering applications like fluid dynamics and materials design.

Physics-Informed Neural Networks with Hard Constraints for Inverse Design

The paper presented by Lu et al. explores the application of Physics-Informed Neural Networks (PINNs) with hard constraints specifically in the context of inverse design problems. The research intricately integrates neural network architectures with physical laws in the form of partial differential equations (PDEs), catering particularly to electromagnetic holography problems. This approach seeks to amalgamate the strengths of data-driven models with physically consistent solutions, minimizing the reliance on vast datasets typical in traditional machine learning paradigms.

Methodological Framework and Numerical Insights

The authors employ challenging constraints derived from Maxwell's equations, specifically focusing on the holography problem represented through the electromagnetic wave equation. The constraint is applied directly to the neural network's training process, ensuring adherence to physical laws. The electromagnetic fields are assumed to be time-harmonic, with specific boundary conditions applied at material interfaces. Of note, the paper exemplifies this approach by demonstrating a finite-difference frequency-domain method to optimize permittivity within a constrained design space.

For practical computation, the study employs two methods: the finite-difference frequency-domain (FDFD) approach and the finite element method (FEM). In the FDFD method, a pixel resolution considerably higher than the wavelength in vacuum confines the permittivity design within predefined bounds, and the optimization problem is resolved using the method of moving asymptotes (MMA). Notably, convergence is achieved within 5000 steps, indicating the computational robustness of the method. The FEM approach uses adaptive mesh resolution with critical topology optimization techniques.

Significantly, the paper also incorporates a smoothed profile method for simulating Stokes flow, highlighting its multi-faceted computational applications. The method articulates the interaction between solid and fluid domains using a hyperbolic tangent function and captures flow behavior using the Stokes equations with appropriate boundary conditions.

Implications and Future Directions

Theoretically, the incorporation of hard constraints within PINNs presents a promising avenue for integrating deep learning with physical simulations, potentially pioneering a paradigm shift in AI-driven PDE solvers. This approach is particularly potent for solving complex inverse design problems where explicit solutions are challenging, if not infeasible. Practically, this methodology could substantially enhance the efficiency and accuracy of computational designs across diverse engineering disciplines.

Future exploration could extend this methodology to broader classes of PDEs and constraint types, promoting its adoption in real-world physics applications such as fluid dynamics, structural mechanics, and materials design. Furthermore, optimizing network architectures and hyperparameters in the context of hard-constrained PINNs could prove beneficial in refining performance metrics and reducing computational overhead. The intersection of PINNs with advanced computational techniques continues to be rich with potential, warranting further empirical and theoretical investigation.