💻 computer science

Residual Physics-Informed Neural Networks for High-Fidelity BLDC Motor Modeling

This paper introduces a Residual Physics-Informed Neural Network (PINN) that rapidly learns a continuous-time surrogate for six-state BLDC motor dynamics, achieving real-time inference speeds up to 118 times faster than conventional ODE solvers while maintaining high fidelity through a curriculum-scheduled physics-data loss.

Original authors: Haitham El-Hussieny

Published 2026-07-13

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Haitham El-Hussieny

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you're trying to teach a robot how to move its legs perfectly. To do that, the robot needs to understand exactly how its motors work. Usually, engineers use two ways to figure this out: they either write down complex math equations (like a strict recipe) or they train a "black box" computer brain (a neural network) to guess the answers based on data.

The problem? The math recipe is hard to get right if you don't know every tiny detail, and the black box brain often makes up wild guesses that break the laws of physics.

This paper introduces a new kind of "smart brain" called a Residual Physics-Informed Neural Network (ResNet-PINN). Think of it as a student who is taking a test but is also forced to check their work against a rulebook while they write. Instead of just memorizing answers, this student learns the rules of the game (the physics) so they can't cheat.

The "Rulebook" and the "Student"
The motor in question is a Brushless DC (BLDC) motor, the kind used in high-tech robots. It has six things happening at once: where the rotor is, how fast it's spinning, the current in three different wires, and how hot the motor is getting.

The authors built a tiny, compact neural network (a student with only about 1,300 "neurons" or parameters) to predict all six of these things at once. But here's the magic trick: they didn't just let the student guess. They gave the student a "physics penalty." If the student's guess violates the laws of electricity or heat, the teacher (the training algorithm) gives them a big red mark.

The "Warm-Up" Trick
There was a snag, though. If you force a student to follow strict rules before they even know the basics, they get confused and stop learning. The authors solved this with a "curriculum scheduling" strategy. Imagine a teacher who says, "For the first 50 minutes, just guess the answers. Then, for the next 350 minutes, I'll slowly start checking your work against the rulebook. After that, I'll be strict." This "warm-up" allowed the network to learn the data first, then slowly learn the physics without getting overwhelmed.

How Fast is It?
The real superpower of this model is speed.

The Old Way: Using traditional math solvers to calculate the motor's movement is like walking through a maze one step at a time. It's accurate, but slow.
The New Way: This PINN is like a teleporter. Once trained, it can predict the motor's state in 0.1 to 22 microseconds (that's 0.0000001 to 0.000022 seconds).

In tests, this new method was up to 118 times faster than the standard math solvers when checking many points at once. It's so fast that it could theoretically run inside a real-time robot controller, helping the robot react instantly.

What It Got Right (and What It Didn't)
In simulations, this "smart brain" was very good at predicting the slow-moving parts, like the motor's angle and temperature, with an error of less than 14%. However, for the fast-moving electrical currents, the error was a bit higher, around 31%. The authors suggest that if they made the network bigger (more neurons), they could likely get the error down to under 5% while still keeping it incredibly fast.

What It's NOT
It's important to note that this paper is based entirely on simulations. The authors built a virtual motor and trained the AI on that. They have not yet tested this on a real, physical motor in a lab, nor have they proven it works on actual hardware yet. They also explicitly ruled out using "Hall-sensor commutation" (a specific type of switching mechanism) in their model; they focused only on the smooth, sinusoidal wave patterns.

The Bottom Line
This paper suggests that by combining a tiny, efficient neural network with a "rule-checking" system and a smart learning schedule, we can create a digital twin of a motor that is both physically accurate and blazingly fast. While it's currently a simulation, it points the way toward robots that can think and react in the blink of an eye, provided we can eventually prove it works on real metal and wires.

Technical Summary: Residual Physics-Informed Neural Networks for High-Fidelity BLDC Motor Modeling

Problem Statement
Accurate dynamics modeling of Brushless DC (BLDC) motors is essential for high-performance control in robotic joints and aerial vehicles. While classical white-box models based on first-principles Ordinary Differential Equations (ODEs) are interpretable, they require precise parameter identification; even minor errors in resistance or inductance can degrade control performance. Conversely, black-box neural networks can approximate complex mappings but often violate physical laws when extrapolating beyond training distributions. Furthermore, standard numerical ODE solvers (e.g., Runge-Kutta) struggle to meet the strict latency requirements (<1 ms) of real-time controllers, particularly when high accuracy is required.

Methodology
The authors propose a Physics-Informed Neural Network (PINN) with a deep residual (ResNet) backbone to create a continuous-time surrogate model for a three-phase BLDC motor.

Architecture: The model utilizes a ResNet architecture with two residual blocks and a hidden dimension of 16, resulting in approximately 1,300 trainable parameters. This compact design is chosen specifically for real-time deployment. The network maps inputs of time ( $t$ ), excitation frequency ( $f$ ), amplitude ( $A$ ), and applied three-phase voltages ( $V_a, V_b, V_c$ ) directly to the six-dimensional state vector: rotor angle ( $\theta$ ), angular velocity ( $\omega$ ), three-phase currents ( $I_a, I_b, I_c$ ), and winding temperature ( $T$ ).
Physics Constraints: The training loss is a composite of data loss (Normalized Mean Squared Error against simulation data) and a physics loss. The physics loss is derived from the residuals of the governing electromechanical and thermal ODEs. Crucially, these residuals are computed via automatic differentiation of the network output with respect to time, ensuring gradients flow back through the network weights to enforce physical consistency.
Training Strategy: To address the difficulty of optimizing stiff ODEs, the authors employ a curriculum scheduling strategy. The physics penalty weight ( $\lambda_p$ ) is zero during an initial warm-up phase (epochs 1–50) to allow the network to fit the data, then linearly ramped up to 1.0 by epoch 400. This prevents the physics constraints from disrupting early data fitting.
Normalization: The ODE residuals are normalized by characteristic scale vectors derived from motor parameters to ensure balanced gradient contributions across different state variables.

Key Contributions

ResNet-PINN Surrogate: Development of a continuous-time surrogate mapping excitation parameters to the full six-state BLDC dynamics, utilizing a ResNet backbone to mitigate vanishing gradients and spectral bias common in deep PINNs.
Curriculum Scheduling & Composite Loss: Implementation of a per-channel ODE residual normalization and a dynamic curriculum schedule that gradually activates the physics penalty, ensuring stable convergence.
Systematic Latency Benchmarking: A rigorous comparison of PINN inference speeds against explicit-Euler and Runge-Kutta (RK45) solvers across batch sizes ranging from 1 to 10,000.

Results

Accuracy: On a test trajectory with unseen excitation parameters, the compact model achieved Normalized Root Mean Square Errors (NRMSE) of <14% for slow states (angle and temperature) and <32% for fast electrical currents. The authors note that the current model prioritizes inference speed; scaling the architecture to larger dimensions is expected to reduce NRMSE below 5%.
Training Efficiency: Training was completed in approximately 110 seconds on a standard CPU using a mini-batch size of 1,024.
Inference Speed: The PINN demonstrated significant speed advantages over traditional solvers. For a single query, the PINN was 14.2× faster than the Euler solver and 10.8× faster than RK45. As batch size increased to 10,000 points, the PINN's throughput grew linearly due to parallelization, achieving latencies of 0.1–22 µs per query. This represents a speed-up of up to 118× compared to the RK45 solver.

Significance and Claims
The paper claims that the proposed ResNet-PINN successfully bridges the gap between data-fitting accuracy and physical interpretability for BLDC motor dynamics. By embedding differential equation residuals directly into the training loss, the model produces a physically consistent surrogate without relying exclusively on dense finite-element simulations.

The primary significance lies in the inference speed. The authors assert that the PINN's ability to process large batches with microsecond-level latency makes it suitable for real-time observer and control applications, such as Model Predictive Control (MPC), where conventional ODE solvers are computationally prohibitive. The work establishes that with proper architectural choices (ResNet) and training strategies (curriculum learning), PINNs can serve as practical, high-fidelity surrogates for real-time mechatronic control.

More like this