High-Precision Ground Characterization of Test-Mass Magnetic Properties for the Taiji Gravitational Wave Mission via a Physics-Informed Neural Framework

This paper proposes an AI-enhanced Differentiable Weighted Least Squares (AI-WLS) framework that combines a dilated residual network with a physical solver to achieve high-precision characterization of test-mass magnetic properties by effectively suppressing non-stationary noise in torsion-pendulum measurements for the Taiji gravitational wave mission.

The Gist

The Mission: Listening to the Universe’s "Whispers"

Imagine you are trying to listen to a tiny, delicate whisper in the middle of a roaring heavy metal concert. That whisper is a gravitational wave—a ripple in the fabric of space-time caused by massive cosmic events like black holes colliding.

To hear these whispers, scientists are building a space mission called Taiji. It involves three spacecraft floating in space, acting like a giant ear. But there is a catch: to hear the universe, the "inner ear" of these spacecraft (called test masses) must be incredibly still. If they wiggle even a tiny bit because of local magnetic fields or tiny bumps, the "whisper" of the universe is lost in the noise.

The Problem: The "Noisy Room" Problem

Before we launch Taiji into space, we have to test these test masses on Earth using a device called a torsion pendulum (essentially a super-sensitive hanging scale that measures tiny twists).

However, testing on Earth is like trying to perform surgery in a windstorm. The laboratory isn't perfectly quiet. There are "glitches"—maybe a truck drives by, an air conditioner kicks on, or the building slightly tilts. These create "colored noise"—unpredictable, messy, and non-stop interference.

Scientists usually use two old-school math tools to clean up this noise:

1. OLS (The Average Joe): It tries to find the middle ground of all the data. But if a giant "glitch" happens, the Average Joe gets distracted and gives you a wrong answer.
2. KF (The Predictor): It tries to guess what happens next based on what happened before. But if the noise changes suddenly (like a sudden gust of wind), the Predictor gets confused and starts making bad guesses.

In short: The old math tools are being tricked by the messy environment.

The Solution: The "Smart Filter" (AI-WLS)

The researchers in this paper created a new, high-tech way to clean the data. They call it AI-WLS. Think of it as a Smart Noise-Canceling Headphone that is also a Master Detective.

Here is how it works using three parts:

1. The Detective (The AI Neural Network):
Instead of assuming the noise is always the same, they built an AI "Detective." This AI scans the incoming data stream in real-time. Its only job is to look at every single millisecond of data and ask: "Is this a real signal, or is this just a glitch from the air conditioner?" If it sees a glitch, it marks that moment as "untrustworthy."

2. The Weighted Scale (The Differentiable Solver):
In the old way, every piece of data was treated as equally important. In the new way, the AI gives every piece of data a "Confidence Score" (a weight).

• Clean data gets a high score (like a clear, loud note in a song).
• Glitchy data gets a very low score (like a sudden loud bang).
  When the math is done, the system listens mostly to the high-score data and mostly ignores the "garbage" data.

3. The Physics Teacher (The Differentiable Framework):
This is the most clever part. Usually, AI is a "black box"—it gives you an answer, but you don't know why. Here, the AI is tethered to the Laws of Physics. The system knows exactly how magnetism should behave. If the AI tries to "cheat" by ignoring real data to make the noise look smaller, the Physics Teacher steps in and says, "No, that doesn't follow the laws of magnetism!" This forces the AI to learn how to be a better detective without breaking the rules of science.

The Result: Crystal Clear Precision

When they tested this "Smart Filter" against the real, messy noise from their laboratory, the results were stunning:

• The Old Tools (OLS and KF): They failed. They couldn't meet the strict requirements needed for the Taiji mission. They were too easily fooled by the "wind" in the room.
• The New AI Tool (AI-WLS): It passed with flying colors. It was able to see through the mess and measure the magnetic properties of the test masses with incredible accuracy—meeting all the requirements for the space mission.

Why does this matter?

By perfecting this "Smart Filter" on Earth, we ensure that when Taiji finally reaches space, its "ears" will be tuned perfectly. We won't be distracted by tiny magnetic wobbles, allowing us to hear the most profound, ancient secrets of the universe.

Technical Summary

Technical Summary: High-Precision Ground Characterization of Test-Mass Magnetic Properties for the Taiji Mission

1. Problem Statement

The Taiji mission, a space-borne gravitational wave observatory, requires ultra-stable Gravitational Reference Sensors (GRS). A critical source of residual acceleration noise is the coupling between the test mass (TM) magnetic properties—specifically its remanent magnetic moment ($\vec{m}_r$) and volume magnetic susceptibility ($\chi$)—and the fluctuating environmental magnetic fields.

To ensure mission success, these parameters must be characterized on the ground with extreme precision (targeting $|\vec{m}_r| < 1 \times 10^{-9} \text{ A}\cdot\text{m}^2$ and $\chi < 1 \times 10^{-7}$). However, traditional characterization using torsion pendulums faces a major obstacle: the experimental background is non-stationary and "colored" (containing low-frequency drifts and transient glitches). Standard estimation methods, such as Ordinary Least Squares (OLS) and the Kalman Filter (KF), assume stationary Gaussian noise. In realistic environments, these methods introduce systematic biases that exceed the stringent Taiji mission requirements.

2. Methodology: The AI-WLS Framework

The authors propose an AI-enhanced Differentiable Weighted Least Squares (AI-WLS) framework. This approach fuses deep learning with analytical physics to create an end-to-end differentiable pipeline. The framework consists of three integrated blocks:

• Block I: Physics-Fixed Forward Operator: This module uses the known geometry of the magnetic coils and the driving current waveforms to construct a design matrix $\mathbf{A}$. It preserves the exact linear physical mapping from magnetic parameters to the expected torque response.
• Block II: Learned Noise Evaluator (Dilated ResNet): Instead of using a fixed noise model, a 1D Residual Network with exponentially dilated convolutions and Squeeze-and-Excitation (SE) modules acts as a dynamic noise evaluator. It analyzes the raw torque time series to identify "contaminated" segments (glitches/drifts) and assigns a time-varying, strictly positive weight $w_n$ to each sample via a Softplus activation.
• Block III: Differentiable WLS Solver: The weights from the neural network are fed into a closed-form Weighted Least Squares solver. Because the entire pipeline (from weights to parameter estimation) is differentiable, the error in the estimated magnetic parameters can be backpropagated through the solver to train the neural network.

Training Strategy: The network is trained by superposing simulated magnetic signals onto real "free-run" noise recorded from the CIOMP torsion pendulum. A Weighted Mean Absolute Percentage Error (WMAPE) loss function is used to ensure that all parameters (which differ by orders of magnitude) are constrained equally.

3. Key Contributions

• Hybrid Architecture: Successfully combines the interpretability and precision of analytical physics with the adaptive power of deep learning.
• End-to-End Differentiability: Unlike "black-box" regression, the AI-WLS learns weights specifically optimized to minimize the error of the physical parameters, not just to denoise the signal.
• Robustness to Non-Stationarity: The framework explicitly addresses the failure modes of classical estimators by treating noise as a sample-local, time-varying phenomenon.

4. Results

The framework was validated using real noise data from the Changchun Institute of Optics, Fine Mechanics and Physics (CIOMP) torsion pendulum. The performance was benchmarked against OLS and KF across 18 experimental scenarios (varying driving currents and coil configurations).

• Precision Achievement: AI-WLS was the only method that strictly satisfied all Taiji ground-test requirements. It bounded the maximum absolute errors at:
  • $|\Delta \vec{m}_r| \leq 4.46 \times 10^{-10} \text{ A}\cdot\text{m}^2$
  • $|\Delta \chi| \leq 7.8 \times 10^{-8}$
• Comparison with Baselines:
  • OLS failed to meet targets due to its inability to down-weight drift-dominated segments.
  • KF exhibited persistent systematic bias because its stationary covariance model could not distinguish between slow environmental drifts and the actual magnetic signal.
• Sensitivity Improvement: Compared to existing literature (Yin et al.), the AI-WLS framework improved precision by approximately one to two orders of magnitude across all parameters.

5. Significance

This work provides a methodological foundation for the pre-launch verification of the Taiji mission. Beyond magnetism, the AI-WLS framework is a general-purpose tool applicable to any ultra-precision measurement task where a known linear physical model must be extracted from non-stationary, non-Gaussian noise. This includes characterizing other stray-force couplings (thermal, electrostatic, etc.) for the Taiji and LISA-like space-borne gravitational wave missions.