Dual-Network PINNs for Optimal Control: A Reproducible Benchmark on the Mass-Spring-Damper System

2026-06-13 · Source: Machine Learning · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Advanced, quick

Summary

A reproducible benchmark study introduces a direct dual-network Physics-Informed Neural Network (PINN) formulation for optimal control of a mass-spring-damper system. This approach solves the classical linear-quadratic optimal control problem, comparing its performance against Pontryagin's Minimum Principle with single shooting and direct transcription via trapezoidal collocation. The PINN employs two feedforward neural networks: a state network enforcing boundary conditions exactly using a composite cubic-and-mask ansatz, and an unconstrained control network. Its composite loss combines physics residuals with a trapezoidal cost functional approximation. The PINN accurately reproduces the classical optimal cost to four significant digits and satisfies terminal state constraints by construction, with errors comparable to classical methods. While training is approximately two orders of magnitude slower than classical shooting, the work prioritizes methodological clarity and provides a Google Colab implementation to simplify entry into PINN-based optimal control for practitioners.

Key takeaway

For machine learning engineers or AI scientists exploring optimal control, this dual-network PINN formulation provides a transparent and reproducible method. You can achieve high accuracy, matching classical optimal costs to four significant digits, even if training is slower. Utilize the provided Google Colab implementation to practically apply PINN-based optimal control without needing prior expertise in adjoint methods or two-point boundary value problems. This lowers the entry barrier for adopting advanced control techniques.

Key insights

A dual-network PINN formulation achieves high accuracy in optimal control by exactly enforcing boundary conditions, offering a clear, reproducible alternative to classical methods.

Principles

• Dual-network PINNs can solve optimal control problems accurately.
• Exact boundary condition enforcement is possible via composite ansatz.
• Reproducibility and clarity are key contributions.

Method

Optimal control is recast as a constrained optimization problem solved by two feedforward neural networks: a state network with exact boundary conditions via a composite cubic-and-mask ansatz, and an unconstrained control network. The loss combines physics residuals and a trapezoidal cost approximation.

In practice

• Use the Google Colab implementation for PINN optimal control.
• Apply dual-network PINNs to systems like mass-spring-dampers.

Topics

• Physics-Informed Neural Networks
• Optimal Control
• Mass-Spring-Damper System
• Constrained Optimization
• Neural Networks
• Reproducible Benchmarking

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Student

Related on AIssential

• Physics-Informed Neural Network Modeling of Biodegradable Contaminant Transport through GCL/SL Composite Liners — Hard-constrained PINNs significantly improve accuracy and stability for modeling contaminant transport in composite liners.
• Co-Learning Port-Hamiltonian Systems and Optimal Energy-Shaping Control — Co-learning pH system models and optimal EB-PBC policies via alternating optimization ensures stability and robustness in control.
• PINN is predicting trivial solution for stiff ODE [D] — PINNs struggle with stiff ODEs, often predicting trivial solutions beyond certain stiffness thresholds.
• A Wachspress-based transfinite formulation for exactly enforcing Dirichlet boundary conditions on convex polygonal domains in physics-informed neural networks — Wachspress-based transfinite interpolation enables exact Dirichlet boundary condition enforcement in PINNs on complex geometries.
• Adaptive Domain Decomposition Physics-Informed Neural Networks for Traffic State Estimation with Sparse Sensor Data — ADD-PINN improves traffic state estimation from sparse sensor data by adaptively decomposing domains based on residual errors.
• Mesh Graph Neural Network Framework for Accelerating Finite Element Simulation for Arbitrary Geometries — Mesh Graph Networks effectively generalize finite element analysis predictions across arbitrary structural geometries by encoding relative features.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning.