✨ TL;DR
DiLaR-PINN is a physics-informed neural network that learns unmodeled dissipative effects in electromechanical systems by constraining residual terms to be energy-dissipating rather than energy-injecting. The method combines first-principles models with data-driven components that operate on latent states and guarantee physically consistent energy behavior.
First-principles models of electromechanical systems often fail to accurately capture complex dissipative phenomena like joint friction, stray losses, and structural damping. While residual-learning physics-informed neural networks can augment imperfect physical models with data-driven components, existing approaches typically use unconstrained multilayer perceptrons for the residual terms. These unconstrained networks can inadvertently inject artificial energy into the system, violating fundamental physical principles and leading to unrealistic dynamics, especially during long-horizon predictions. This is particularly problematic for embodied systems where accurate dynamical modeling is critical for both simulation and control applications.
The authors propose DiLaR-PINN (Dissipative Latent Residual Physics-Informed Neural Network), which structurally constrains the residual network to guarantee energy dissipation. The key design choices include: (1) the residual network operates only on unmeasurable (latent) state components rather than all states, (2) the network is parameterized in a skew-dissipative form that mathematically guarantees non-increasing energy regardless of the learned parameters, and (3) a recurrent rollout training scheme with curriculum-based sequence length extension is employed to enable stable and data-efficient learning under partial state observability. This architecture ensures that the learned residual terms can only remove energy from the system, never add it, making the augmented model physically consistent with dissipative dynamics.
What the paper shows.
DiLaR-PINN was validated on a real-world helicopter system and compared against four baselines: a pure physical model, an unstructured residual MLP, a soft-constrained DiLaR variant, and a black-box LSTM. The results demonstrate that DiLaR-PINN achieves superior long-horizon extrapolation performance compared to all baselines. The hard dissipative constraint outperforms the soft constraint variant, confirming that structural guarantees are more effective than penalty-based approaches. DiLaR-PINN more accurately captures the dissipative effects present in the real system while maintaining physical consistency, showing that the combination of latent residual learning and guaranteed energy dissipation leads to both better accuracy and more reliable extrapolation.
The paper does not explicitly discuss computational costs or training time compared to simpler baseline methods. The approach requires partial knowledge of the system's physical structure to formulate the first-principles component, limiting applicability to systems where such models are unavailable. The validation is performed on a single real-world system (helicopter), so generalization to other types of electromechanical systems remains to be demonstrated. The method assumes that unmodeled effects are primarily dissipative, which may not hold for all modeling errors or system configurations. Additionally, the curriculum learning strategy introduces hyperparameters for sequence length scheduling that may require tuning for different applications.
✨ Generated by Claude · Apr 21, 2026 · Read the PDF for authoritative content.